Medical Image Format FAQ - Part 3

Proprietary Formats

Access to other parts ...


3. Proprietary Formats

3.1 Proprietary Formats - General Information

3.1.1 SPI (Standard Product Interconnect)

Used for files exported from:

SPI is a standard based on the old ACR/NEMA 1 standard, devised I gather by Siemens and Philips, for use in a PACS environment. Who currently maintains it and whether or not Sienet PACS systems are based on it, I am not certain. Many machines in the workplace use it in some shape or form, or can export files in SPI format. I gather it has been around since 1987 or so, but I do not yet have access to the reference documents, nor permission to disclose their contents, so much of the following is guess work or hearsay from Usenet.

Like the ACR/NEMA standard, SPI is designed to define interconnections between pieces of equipment from the physical level through to the application level. Where appropriate it utilized relevant parts of ACR/NEMA. Unlike ACR/NEMA, I gather that SPI is aware of the concept of networks, objects containing information, the need to uniquely identify instances of objects, and defines an offline file format. Thus in many ways it sounds like the missing link between ACR/NEMA 2.0 and DICOM 3.0.

SPI makes use of ACR/NEMA data elements and groups, and in addition provides "shadow" private odd-numbered groups as dictated by the ACR/NEMA standard for the purpose of storing additional items of information, including a means of uniquely identifying objects, as well as allowing for enumerated values for elements beyond those defined by ACR/NEMA. SPI also defines a byte order for offline storage of data streams. Integers are stored in little endian format (least significant byte first).

The private groups mechanism works as follows. For each odd numbered group (other than 0x0001,0x0003,0x0005,0x0007 and 0xffff), the elements 0x00nn in the range 0x0010 through 0x00ff contain a single valued string identification code that identifies the creator of the range of elements 0xnn00 through 0xnnff. Neat eh ? For example:

	(0x0009,0x0010) PrivateCreatorDataElement 
	(0x0009,0x0011) PrivateCreatorDataElement 
	(0x0009,0x1000) DavidElement1 <...>
	(0x0009,0x1001) DavidElement2 <...>
	(0x0009,0x1100) HarryElement1 <...>
	(0x0009,0x1101) HarryElement2 <...>

You get the idea. The nice thing about this scheme is that each creator dictionary considers its elements numbered from 0x0000, but these will be remapped to a block of elements depending on exactly which PrivateCreatorDataElement is used in the particular data set. Hence multiple groups from different creators can co-exist happily in the same data set, and vary in position between data sets.

Note that the group number IS taken into consideration ... a private element with the same element offset and the same creator will have a different meaning depending on which group it is in.

SPI uses this concept extensively and defines a large dictionary with different creators with convoluted names for different modalities and PACS operations. A few sample elements are described here. Particularly important are those elements for purposes that were not envisaged when ACR/NEMA 1 was written, but are necessary to create valid DICOM 3 data sets. Such things as FlipAngle for MR scans for example. Note that the SPI UID is not the same as a DICOM UID, but presumably it is unique ! Note also that the creator of "SPI RELEASE 1" is the same as "SPI Release 1" and "SPI" ... presumably someone messed up between machines or modalities or manufacturers. For a more extensive SPI data dictionary see the DICOM conversion tools. The value representation fields are shown here using the modern DICOM equivalents rather than the older, less specific ACR/NEMA names. The "owner" is what is used as the string value of the PrivateCreatorDataElement when a range of elements in a group is claimed.

Element     Owner                 Name                  VR VM

(0009,0010) SPI                   Comments              LO  1
(0009,0015) SPI                   UID                   LO  1
(0009,0010) SIEMENS MED           RecognitionCode       LO  1
(0011,0010) SPI RELEASE 1         Organ                 LO  1
(0011,0015) SPI RELEASE 1         AllergyIndication     LO  1
(0011,0020) SPI RELEASE 1         Pregnancy             LO  1
(0011,0010) SIEMENS CM VA0  CMS   RegistrationDate      DA  1
(0011,0011) SIEMENS CM VA0  CMS   RegistrationTime      TM  1
(0011,0023) SIEMENS CM VA0  CMS   UsedPatientWeight     IS  1
(0013,0020) SIEMENS CM VA0  CMS   PatientName           LO  1
(0013,0022) SIEMENS CM VA0  CMS   PatientId             LO  1
(0013,0030) SIEMENS CM VA0  CMS   PatientBirthdate      LO  1
(0013,0031) SIEMENS CM VA0  CMS   PatientWeight         DS  1
(0013,0035) SIEMENS CM VA0  CMS   PatientSex            LO  1
(0013,0040) SIEMENS CM VA0  CMS   ProcedureDescription  LO  1
(0013,0042) SIEMENS CM VA0  CMS   RestDirection         LO  1
(0013,0044) SIEMENS CM VA0  CMS   PatientPosition       LO  1

(0019,0010) SIEMENS CM VA0  CMS   NetFrequency          DS  1
(0019,0011) SIEMENS CM VA0  ACQU  SequenceFileName      LO  1
(0019,0021) SIEMENS CT VA0  GEN   Exposure              DS  1
(0019,0026) SIEMENS CT VA0  GEN   GeneratorVoltage      DS  1
(0019,0050) SIEMENS MR VA0  GEN   NumberOfAverages      IS  1
(0019,0060) SIEMENS MR VA0  GEN   FlipAngle             DS  1
(0019,0012) SIEMENS MR VA0  COAD  MagneticFieldStrength DS  1

(0021,0010) SIEMENS MED           Zoom                  DS  1
(0021,0011) SIEMENS MED           Target                DS  2
(0021,0020) SIEMENS CM VA0  CMS   FoV                   DS  2
(0021,0060) SIEMENS CM VA0  CMS   ImagePosition         DS  3
(0021,0061) SIEMENS CM VA0  CMS   ImageNormal           DS  3
(0021,006a) SIEMENS CM VA0  CMS   ImageRow              DS  3
(0021,006b) SIEMENS CM VA0  CMS   ImageColumn           DS  3
(0021,0039) SIEMENS MR VA0  GEN   SlabThickness         DS  1
(0021,0070) SIEMENS MR VA0  GEN   NumberOfEchoes        IS  1

3.1.2 Siemens - Features common to multiple families

The Numaris (MRI) and Somaris (CT) software contains certain common features, especially when running on common platforms. This is particularly true of more recent versions that are Sparc and SunOS based rather than the older Vax/VMS systems . Siemens Vax/VMS

Under construction. Siemens Sparc SunOS

This information is derived mostly from some recent experiments with Numaris VB21B on an Open and Somaris on an AR-C. There is a lot of useful information to be found in the System Manual for both families, not to mention the configuration release notes. Both use bog standard Sun OS 4.1.x, and tend to keep the platform/application specific information in the /usr/appl tree. The user interface is standard OpenWindows. Starting up

This will become apparent when the system is started up. The normal SunOS boot procedure is observed. On somaris, the system automatically loads Open Windows and followed by the Somaris application. On Numaris one logs in as the "mr" user, usually without any password, and gets OpenWindows and the Numaris application. Interrupting this process will be described later. Getting a console

The first step in exploring the system is getting a console. On Numaris this is easy. Running all the way down the right hand side of the screen is an information area from the Numaris application. About a third of the way down the edge, a little grayed out icon is visible. Clicking or dragging on this will expose the fact that this is an iconified console window. On Somaris, the console is still iconified but completely hidden by the right information area. The trick to grabbing this is to do a System/End (menu with right mouse button down) and select Application and Restart, which brings the application and the OpenWindows down and back up again. While this is happening you can see the iconified console and drag it into the middle of the screen, where you can open it later.

While on the subject of System/End, the various options are permuations of normal commands like logout, halt or shutdown.

Once one has a Unix prompt one can explore the system, and create directories in which to save exported images. The Numaris manual's example suggested /usr/appl/external as a place to store exported files. On Numaris this already exists and is empty. On Somaris it doesn't but the normal user has the permission to create it with a "mkdir /usr/appl/external". The normal commands like telnet and ftp are available if one wants to use these to go outward bound on the network, if it is configured (which will be discussed later). Native images

Images are stored in native form in /usr/appl/data/disk1, at least on the systems that I have examined. They are stored one image per file, and named something like nnn-ss-iii.ima, where nnn is some sequential number that pertains to the patient (or instance of the examination ... I am not sure), ss is the series number (always 1 on Somaris), and iii is the sequential image number within nnn. The hard part is figuring out what nnn is for the patient you want ... this number is not displayed in the normal Patient Select dialogs or anywhere else I can find. Counting back from the latest patient and comparing the highest value of iii seems to be a crude but effective approach.

The native images are stored in the usual Siemens style, with a binary header of fixed length (that varies from product to product in length and layout) and trailing uncompressed image pixel data. The specifics where known are described elsewhere. Exporting images

On any of these products one can use the System/External Data menu option to bring up a dialog with Import or Export choices. Select Export, enter /usr/appl/external or whatever as the destination, and choose the image numbers (eg. "1-6,10,22-24" is quite acceptable) and they will be written where you asked. The patient name must be exactly as it is registered. The catch is that the exported SPI files will be named with the patient's name and the current date and time of export, not the time of acquisition or reconstruction or whatever, so sorting through these to determine what they are is a pain. The form of the date and time stamp in the name is "yyyymmddhhmmssff". Physical connection

So you know where the images are ... how do you get them off. One way is by ethernet connection. One doesn't have to have the PACSNet or DICOM option to be able to connect to the network. If you haven't paid for the PAL that provides hardware protection for these functions, it doesn't mean that the ethernet software in Sun OS and the ethernet port on the Sparc host is not live. During installation of the Somaris or Numaris software the Siemens Field Engineer can configure the interface with a IP address of your choice (it defaults to under Numaris, and the le0 interface is not configured by default under Numaris).

If the Siemens FE is unfamiliar with the procedure tell them to use the "install" login, choose SSC (Site Specific Configuration) then RC (Reset host Configuration), accepting the defaults until you get to "Internet Address". If you know the "install" password (or can change it as root) you can do this yourself. I don't think the additional layer of Siemens password protection applies to this particular tool, though there are many you won't be able to run.

If you are really desperate you can gain root access and manually configure the SunOS network configuration without using the Siemens tool, but you need to be pretty familiar with SunOS to do this. You need to put in a real IP address in /etc/hosts, create an /etc/hostname.le0, and if necessary set up /etc/netmasks if the default is not appropriate. I tried this and it works but it somehow messed up camera communications, so doing it with the Siemens FE is probably better. Don't forget to back up the critical files first just in case.

The standalone configuration on the AR/C had just the loopback address ( in /etc/hosts and no /etc/hostname.le0.

The physical ethernet connector is normally unused, and is located on the Sparc host board and is the usual AUI connector (ie. you need an AUI to 10BaseT or whatever transceiver). On the AR/C I tested it was located under the desktop (ie. lift the desktop off, and then the metal cover), sticking up on the left hand side. On the Magnetom Open it was in the computer room in the cabinet with the host processor at the bottom on the left hand side. In this installation it was connected to a lead going to a breakout panel on the top cover of the cabinet. This is unused so just disconnect it and plugin your own. Archive devices

Another way to get the images off is to just use the QIC streaming tape drive. This is probably still installed in older machines, but the newer software is being distributed on CD-ROM so the tape drive is being pulled and replaced with a CD. It is probably still in the maintenence closet though and would be easy to swap back in. No configuration is necessary. It is accessed as usual as /dev/rst0 and its rewinding and non-rewinding variants, and one can just tar image files off to it. Very handy. No messing with wierd Pioneer WORM's and MOD's !

The drive is physically located on the front of the processor box in the desk models and in the host processor cabinet beside the optical drive in the computer room in the larger installations.

Speaking of WORM's and MOD's, they are the same unreadable media as used by GE, but of course have a different filesystem. When used as archive devices these are not the standard unix file system, and you will not see any evidence of a mounted device doing a "mount" or "df", even though when you stick one in the drive the application automatically detects it and mounts it. It is said in the release notes that one can actually format and mount one of these as a unix filesystem instead (the MOD at least) but I don't know how to do it, and haven't discovered, not possessing one of them there Pioneer drives to read one on. Becoming root

If you thought you could mess up a perfectly good scanner already, try becoming root. Why would one need to do this ? To manually reconfigure the network, to change passwords for critical logins like install, to create your own login some place clean and safe, etc. Since this is standard SunOS, the usual principles apply ... first try rebooting in single user mode. Do this with a System/End choosing System/Norestart and you will get a boot prompt. Type "b -s" and it comes up in single user mode, allowing you to mess with /etc all you like as root.

If this mode has been password protected (and one can do this by removing "secure" from "console" in /etc/ttytab ... see "man 5 ttytab") then one is not out of luck yet. Now you have to put a SunOS boot disk in the CDROM drive (or plugin an external CDROM drive) and boot SunOS mini-root, then mount /dev/sd0 as /mount and you are in business. (If you don't have a SunOS CDROM then you probably shouldn't be doing this kind of thing in the first place). Reset

If you are messing about in SunOS, periodically the Somatom application will get out of sync with the new reality you have created and will complain that an Init/Reset is necessary ... well, do an Init/Reset. I have forgotten exactly where it is in the menus, whether under System or Measurement. It is documented in the system manual and seems harmless.

3.2 CT - Proprietary Formats

3.2.1 General Electric CT

Now we get to the meaty part. After years of being faced with the problem of either a) hours of detective work, or b) tediously tracking down the name of the responsible person and exercising a non-disclosure agreement, this is now no longer necessary, as General Electric are making their image format description documents freely available. For details see the GEMS image format information contacts section later on. In the meantime, both for historical completeness, educational purposes, and for those who can't wait for document to come in the mail, a summary of the relevant formats and decompression algorithms is provided here. GE CT 9800

References (see the GEMS image format information contacts section): GE CT 9800 Image data

Almost everyone in this field has at some stage encountered the dreaded CT 9800 format. The world is divided into two groups of people ... those who have seen the documents or the critical piece of code in another program or have been given a handy hint, and those who will never figure out the format themselves.

Essentially the format fits into the "block format" described earlier, with pointers to each of the major header components. Rarely, if ever, does one encounter a file that doesn't have the same size blocks in the same place, so most people treat it as a fixed layout. I believe that reformatted images may have another header stored in there, but I have never tested for it.

The data itself is stored in one of two forms depending on whether compression is selected or not during archival. In the uncompressed form, a type of perimeter encoding is used (see later section) in which for an essentially circular object, the outer parts of a rectangular image are discarded (and expected to be filled in with a background pixel value during reconstitution of the image). In the case of the CT9800 then, the image pixel data is interpreted using a map, which contains an entry for each row of the image (either 256, 320 or 512 entries) which specifies the length of the row that is actually stored, centered about the midline of the image. This obviously saves a lot of space.

If compression is selected on one of the later model machines, then a form of Differential Pulse Code Modulation is used, in which advantage is taken of the fact that not all the bits of a 16 bit word are need to store a CT value. I gather only 12 bits of data are actually significant, but one can theoretically represent 15 using this scheme. Essentially, the first 16 bit word is read and used as is. Then another byte is read. If its most significant bit is set, then the remaining 7 bits represent a signed difference value relative to the previous pixel. If its most significant bit is not set, then the difference must have exceeded the range of 7 bits, and hence the next byte is read to complete a valid 16 bit word (15 bits really) which is the actual pixel value. The really neat thing about this scheme is that the same algorithm can be used for compressed or uncompressed data as an uncompressed stream of words will never have the most significant bit set !

The following piece of C++ code pulled out of a CT9800 to DICOM translator will give you the general idea. Note that the perimeter encoding map has already been read in. Note in particular the need to deal with sign extension of the difference value. Also note that the code doesn't handle the first pixel specially because its high bit will not be set.

static void
copy9800image(ifstream& instream,DC3ofstream& outstream,
	      Uint16 resolution,Uint16 *map)
	unsigned i;
	Int16 last_pixel;

	for (i=0; i<resolution; ++i) {
		unsigned line	= map[i];
		unsigned start	= resolution/2-line;
		unsigned end	= start+line*2;
		unsigned j;

		// Pad the first "empty" part of the line ...
		for (j=0; j<start; j++) outstream.write16(0);

		// Copy the middle of the line (compressed or uncompressed)
		while (start<end) {
			unsigned char byte;,1);
			if (!instream) break;
			if (byte & 0x80) {
				signed char delta;
				if (byte & 0x40) {
				else {
		    			delta=byte & 0x3f;
			else {
				last_pixel=byte << 8;,1);
				if (!instream) break;
			outstream.write16((Uint16)last_pixel & 0x0fff);

		// Pad the last "empty" part of the line ...
		for (j=end; j<resolution; j++) outstream.write16(0);

What about the rest of the header information and where is this map stored anyway ? Well, the file is described as a series of 256 by 16 bit word blocks, blocks numbered from 0, words numbered from 1, integers are 16 bit words, as follows:

        block 0 - global header

               word 34   - Int - pointer to global header
               word 35   - Int - pointer to exam header
               word 36   - Int - pointer to image header
               word 37   - Int - pointer to image header2
               word 38   - Int - pointer to image map
               word 39   - Int - pointer to image data
               word 40   - Int - number of blocks in global header
               word 41   - Int - number of blocks in exam header
               word 42   - Int - number of blocks in image header
               word 43   - Int - number of blocks in image header2
               word 44   - Int - number of blocks in image map
               word 45   - Int - number of blocks in image data

Now almost always the layout is as follows, for non-reformatted images:

        block 0 - global header
        block 1 - exam header
        block 2 - image header
        block 3 - image header 2
        block 4 - image map
        block 6 - image data

For reformatted images the layout is said to be different, but I have never seen a description of the contents of the so-called "arrange header", nor do I know where in the global header the pointer and length are stored:

        block 0 - global header
        block 1 - exam header
        block 2 - image header
        block 3 - image header 2
        block 4 - arrange header
        block 9 - image map
        block 11 - image data

Some of the more important contents of the various headers are listed here. For more complete information get the documents from GE or study any one of a number of programs kicking around to dump the header of this kind of file (see sources later). Integers are 16 bit words, ascii strings are Fortran style specifications with two characters per word, and reals are 4 bytes long (see Host machines - Data General):

        block 0 - global header

               word 17-23    - 7A2  - file name

        block 1 - exam header

               word  4       - Int  - exam number
               word  5-11    - 7A2  - exam number
               word  12-17   - 6A2  - patient id
               word  18-32   - 15A2 - patient name

        block 2 - image header

               word  11      - Int  - position (study) number
               word  13      - Int  - group type (2=scout,3=standard,4=dynamic)
               word  14      - Int  - group number
               word  47      - Int  - scan number
               word  48      - Int  - image number
               word  50      - Int  - patient orientation (1=head first,2=feet)
               word  51      - Int  - AP orientation (1=prone,2=sup,3=lt,4=rt)
               word  55      - Int  - contrast (0=no,1=yes)
               word  93-94   - Real - gantry tilt
               word  95-96   - Real - table height mm
               word  97-98   - Real - axial table location mm
               word  124     - Int  - image size (256,320,512) NOT FOR SCOUTS
               word  132     - Int  - detectors/view        - width  for scouts
               word  137     - Int  - compressed views/scan - height for scouts
               word  144-145 - Real - X diameter of recon mm
               word  146-147 - Real - Y diameter of recon mm
               word  155-156 - Real - magnification factor
               word  157-158 - Real - X center
               word  159-160 - Real - Y center
               word  175     - Int  - image map used (1=yes,2=no)
               word  218     - Int  - file type (1=prospective,2=scout,
                                      5=screen save,6=plot)
               word  219     - Int  - data range (number of bits)
               word  236     - Int  - scout orientation (0=ap,1=lateral)
                                      (the 9800 rotates the scout magically)

It is important to check the filetype and image map used entries, particularly if trying to read scouts rather just prospective images. If the map is not in use, it is filled with zeroes and hence if the flag is not checked a simplistic demapping algorithm will fail. Furthermore the number of rows and columns in the image is not specified as such. For prospective images, the imagesize field is valid for both (images are square). For scouts, one must use the detectors/view field for the width and the compressed views/scan field as the height.

The filename entry is quite useful. Therein is stored the RDOS filename of the image, which follows the following convention:

        s     = originating scan station id
        eeeee = exam number
        pp    = prs number (position related set)
        dd    = image number
        tt    = file type
                YP = prospective
                YV = scout
                YR = retrospective
                YG = segmented recon
                YS = screen save
                YL = plot
                YF = reformatted

        eg. B038500165.YP

Having said this, my GE 9800 stores its scouts on tape at least with no file extension at all, rather than the .YV that it is supposed to use. GE CT 9800 Tape format

Probably more CT images have been exchanged for clinical and research purposes using GE 9800 9-track magnetic tapes than any other means. These things are just ubiquitous, particularly considering the proliferation of services providing 3D reconstruction and fabrication a few years ago. Fortunately the format is easy to deal with. The tapes are produced on a primitive DG tape drive and hence are never more than 1600bpi. The first thing on the tape is a directory consisting of two 4096 word (8192 byte) records, then two EOF marks, then 20" of blank tape (because the directory keeps getting updated) followed by image files each separated by an EOF mark and finally an additional EOF mark after the last file.

I won't describe the tape directory format here unless someone specifically asks for it, though it is very simple. I usually just read everything on the tape and sort the files out later. Remember that their filenames are stored in the global header.

Don't forget to set the input magnetic tape record size to 8192 bytes when you are copying these files. If you don't do this some systems quietly truncate each record to some default size. It took me a week to figure out why my files were screwed up the first time I tried this on a DG under AOS/VS (I was desperate and using a networked Signa to read files off a non-networked 9800).

A simple script to read an entire tape from a SCSI tape drive /dev/nrst1 under SunOS, which will peek in each image file to extract the correct filename (simpler than trying to decipher the directory) looks like this:


echo "Rewinding"
mt -f /dev/nrst1 rewind

echo "Extracting directory ..."
dd if=/dev/nrst1 ibs=8192 of=TAPEDIR

while dd if=/dev/nrst1 ibs=8192 of=tape.tmp
	name=`dd if=tape.tmp ibs=16 skip=2 count=1 2>/dev/null`
	if [ -z "$name" ]; then break; fi
	mv tape.tmp $name
	echo "Extracted $name"

echo "Rewinding"
mt -f /dev/nrst1 rewind
echo "Finished" GE CT 9800 Raw data MR

No idea about this one ... I have never had the need or seen any documention. Anyone who does or has please fill in this space. GE CT Advantage - Genesis

References (see the GEMS image format information contacts section):

General Electric now uses the same Sun based architecture for its Advantage CT and Signa 5X MR family, referred to as Genesis, and hence the general details of this scheme will be discussed under the GE MR Signa 5.x - Genesis section. Specifics related to the CT modality will be described here. GE CT Advantage Image data

The Image Extract Tool is used in the same way as on the Signa to extract an image from the database into a single file, either asis or using the requested compression and packing mode. The Genesis file contains headers consisting of several components in common with MR and then a specific CT or MR header. Theroetically, one should be able to use "/usr/g/insite/bin/ximg -g" to extract a prototype C header file describing the file format, as on the Signa, though last time I tried this on a High Speed Advantage this didn't work. Some of the more interesting fields in the CT image header include:

        image header - for CT (1020 bytes long):

                194 - float    - table start Location
                198 - float    - table end Location
                202 - float    - table speed (mm/sec)
                206 - float    - table height
                224 - float    - gantry tilt (degrees) GE CT Advantage Archive format

See the GE MR Signa 5.x Archive format. GE CT Advantage Raw data

Again, unknown. Please fill in this space. GE CT Pace

References (see the GEMS image format information contacts section):

The Pace is a CT scanner made by Yokogawa Medical Systems(YMS) in Japan. The format documents I have for it are partially in Japanese and partially in English, but they get the job done. I have only tested the following on a few images that were extracted off a nine-track tape, so the offsets to the image header fields may not be correct in other cases, but here are "eye-catcher" fields at the start of each header which should be easy to find. The format seems to be shared with the GE MR Max family.

The images described in the documents have a 512 byte study header that begins with "!STD" and an image header of 1024 bytes that begins with "!IMG". In the image that I had to play with, there was a 256 byte header that I am not familiar with tacked on the front, presumambly something to do with being a mag tape rather than a disk image. Anyway this meant that the offset to the study header was 256 bytes, the image header was 768 bytes, and the compressed image data began at 1792 bytes.

I don't know what kind of host is used on the Pace though I have seen some cryptic references to "DOS-68K" in the documents. Regardless, the integers are 16 or 32 bit big-endian. The image data is stored as SIGNED not unsigned 16 bit values, same as on the Sytec and presumably all the YMS systems. Most of the useful dates and times are provided as string values, however there are some dates and times that are 32 bit binary integers. Though not specified in the docs it seems that the dates are days since an epoch of "0 Jan 1980" and the times are in milliseconds. Floats are 32 bit IEEE format, dfined in the Pace documentation as follows:

	bit  31		sign (s)	(0 is +ve)

	bits 30-23	exponent (e)
			- unsigned integer
			- e == 0 for denormalized numbers
			- 0 < e < 255 for normalized numbers
			- e == 255 for other reserved operands

	bits 22-0	significand (f)

	Normalized numbers:
			- bias 127
			- range 0 < e < 255
			- interpreted as 1.f
			- range 1.0 <= f < 2.0

		(-1)^s * 2^(e-127) * 1.f

	Denormalized numbers:
			- e == 0
			- bias 126
			- interpreted as 0.f
			- range f != 0

		(-1)^s * 2^(-126) * 0.f

	Signed Infinities:
		- e == 255
		- f == 0

		- e == 255
		- f != 0

The image header has a chunk in the middle where different values are defined for CT and MR. One can use the first byte of the model number field to distinuish the two modalities. Some of the more important study and image header values are:

    Study header (offset 256 bytes, length 512 bytes):

        Offset  Type    Length  Meaning       Units or values

        0x0     string  4        Eyecatcher         !STD
        0x6     byte    1        Modality           1=CT,2=MR
        0xa     string  5        Study Number
        0x10    datestring       Study Date         yyyy/mm/dd
        0x1a    timestring       Study Time         hh/mm/
        0x26    string  12       Patient ID
        0x36    string  12       Patient Name
        0x50    string  6        Patient Age        yyy;mm
        0x5c    string  2        PatientSex"        'M ','F '
        0xbc    string  4        Contrast media     'NO C','+C  '

    Image header (offset 768 bytes, length 1024 bytes):

        Offset  Type    Length  Meaning       Units or values

        0x0     string  4       Eyecatcher          !IMG
        0x6     byte    1       Modality            1=CT,2=MR
        0xa     string  5       Study Number
        0x10    string  2       Series Number
        0x12    string  2       Acquisition Number
        0x14    string  2       Image Number
        0x20    datestring      Image Date          yyyy/mm/dd
        0x2a    timestring      Image Time          hh/mm/
        0x40    string  2       'H '=Head First,'F '=Feet First
        0x42    string  2       'SP'=Supine,'PR'=Prone,
                                'LL'=Left Lateral Decubitus,
                                'RL'=Right Lateral Decubitus,'OT'=Other
        0x44    string  6       Anatomic location
        0x50    string  4       'AX  '=Axial,'SAG '=Sagittal,'COR '=Coronal
        0x54    float32         Slice position by body coords HF mm
        0x58    float32         Slice position by body coords AP mm
        0x5c    float32         Slice position by body coords LR mm
        0x6c    string  4       Scan fov cm
        0x70    string  4       Scan thickness mm
        0xa0    string  4       Contrast media      'NO C','+C  '

        0x188   float32         Recon center X mm
        0x18c   float32         Recon center Y mm
        0x190   string  4       Recon FOV cm [xx.x]
        0x1a0   u_int16         Pixels in X-axis
        0x1a2   u_int16         Pixels in Y-axis
        0x1a4   float32         Pixel size mm
        0x1b0   float32         Mag center X mm
        0x1b4   float32         Mag center Y mm
        0x1b8   float32         Mag factor

    For CT only:

        0xc8    string  5       Gantry tilt machine coords degrees
        0xe0    string  5       Exposure time ms
        0xe6    string  3       Tube current mA
        0xea    string  5       Exposure mAS
        0xf0    string  3       KVP
        0xf4    string  2       'CW'=Clockwise,'CC'=CounterClockwise

    For MR only:

        0xc0    string  5       Tilt ordered by user Axis+/-Angle [xx+/-xx]
        0x100   string  2       Echo number
        0x102   string  2       Number of echoes
        0x104   string  2       Slice number
        0x106   string  2       Number of slices
        0x108   string  2       Number of excitations
        0x10a   string  5       Repetition time ms
        0x110   string  5       Inversion time ms
        0x115   string  5       Echo time ms
        0x130   string  4       Magnetic flux density (T)

Unlike the Sytec sample images I had, compression was used in the Pace images I received. This is a neat scheme that uses both Run Length Encoding and Differential Pulse Code Modulation. Essentially, each byte may be a flag value 0x81 which indicates the next byte is a run length of the current pixel, or a flag value 0x80 which indicates that the current mode should be toggled between "reference" mode, in which the subsequent 16 bit words are new pixel values, or "difference" mode, in which case subsequent bytes are signed differences added to the current pixel value. The initial mode is "reference" mode. Runs do extended across horizontal line boundaries.

I am not totally clear from the documentation or the sample images where in the header is the flag to say compression is in use or not. It is probably bit 5 of the Image Attribute field in offset 0x1ac in the image header, where a false value specifies DPCM and a true value specifies uncompressed or "Original" encoding. The docs say this is for optical disk only, but the compressed image from tape I have has this bit false, which is correct.

The following piece of code will decode such a compressed image:

static void
copypaceimage(istream& instream,ostream& outstream,
	      Uint16 width,Uint16 height)
// NB. the exclusive or with 0x8000 makes the signed Pace values unsigned
// which is what the PGM convention is ... just omit the ^0x8000
// everywhere if you want the data left signed.

	unsigned i;
	Int16 pixel=0;
	enum Mode { Difference, Reference } mode = Reference;
	for (i=0; i<height*width;) {
		unsigned char byte;,1);
		if (!instream) break;
		if (byte == 0x80) {		// Mode switch
			if (mode == Difference)
		else if (byte == 0x81) {	// Run length flag,1);
			if (!instream) break;
			unsigned repeat=byte;
			while (repeat--) write16little(outstream,pixel^0x8000);
		else {
			if (mode == Difference) {
				pixel+=(signed char)byte;
			else {
				if (!instream) break;
	if (!instream) cerr << "Premature EOF byte " << i << "\n" << flush;
} GE CT Sytec

I don't have one of these either, and it turns out that the format is NOT the same as the Pace as GE Milwaukee initially thought. The format may be shared with the Vectra, but this is not known for certain. I do have a few sample images and have worked out many of the values in the headers. The format may be available from Yokogawa in Japan. Milwaukee apparently doesn't have it.

The host is an MS-DOS clone using the J-DOS operating system, a Japanese version of DOS to handle 16 bit Kanji characters. Alan Rowberg tells me it has a 5.25" drive that writes disks that are unreadable by anything else in the universe.

The images have a header of 3752 bytes and are followed by 16-bit signed integers. The surround is -1500 which is probably -1500 H.U. The sample files I had did not use any form of compression.

The data formats are big-endian. Fortuitously the date/time format is the same as unix ... a 32 bit unsigned integer containing seconds since an epoch of 00:00:00 GMT 1 Jan 1970. Floats are 32 bit IEEE format as described in the Pace format.

The head first/feet first and prone/supine fields in the Sytec file are not known. The sense and identification of corners in the Sytec sample files was done by guess work, and may be wrong if the samples weren't scanned head first supine, and the images are not supposed to be looked at from bottom up in the usual convention.

The header is 3752 bytes long. The known header values are (byte offsets from 0):

      Offset   Type         Meaning               Units or values

        7      string       ModelNumber 

        126    string       Organization
        204    string       PatientID
        217    string       PatientName

        328    datetime     ExamDateTime 
        402    string       ExamDescription
        425    string       Modality
        444    string       ExamStationID

        1164   int16        ExamNumber
        1166   int16        SeriesNumber
        1172   datetime     SeriesDate 
        1176   string       SeriesDescription
        1206   string       SeriesStationID

        1224   int16        ScanType                # 1=axial,3=scout
        1240   string       AnatomicalReference        

        1280   float32      SeriesStartLocation
        1288   float32      SeriesEndLocation

        2192   u_int16      ImageExamNumber 
        2194   u_int16      ImageSeriesNumber 
        2196   u_int16      ImageNumber                
        2204   datetime     ScanDateTime 
        2208   float32      ScanDuration            #? secs
        2212   float32      SliceThickness          # mm
        2216   u_int16      XMatrix
        2218   u_int16      YMatrix        
        2220   float32      FieldOfView             # mm
        2224   float32      ScoutLength             # mm
        2228   float32      XDimension              # mm
        2232   float32      YDimension              # mm
        2236   float32      XPixelSize              # mm
        2240   float32      YPixelSize              # mm

        2310   u_int16      ScoutOrientation        # 0=none,1=ap,2=lateral
        2316   float32      TablePosition           # mm
        2320   float32      SliceCenterX            # mm
        2324   float32      SliceCenterY            # mm
        2328   float32      SliceCenterZ            # mm
        2332   float32      NormalVectorX           # unitized
        2336   float32      NormalVectorY           # unitized
        2340   float32      NormalVectorZ           # unitized
        2344   float32      TopRightHandCornerX     # mm
        2348   float32      TopRightHandCornerY     # mm
        2352   float32      TopRightHandCornerZ     # mm
        2356   float32      TopLeftHandCornerX      # mm
        2360   float32      TopLeftHandCornerY      # mm
        2364   float32      TopLeftHandCornerZ      # mm
        2368   float32      BottomLeftHandCornerX   # mm
        2372   float32      BottomLeftHandCornerY   # mm
        2376   float32      BottomLeftHandCornerZ   # mm
        2384   float32      ScoutStartLocation      # mm
        2388   float32      ScoutEndLocation        # mm
        2408   int32        GeneratorVoltage        # kVP
        2412   int32        TubeCurrent             # mA
        2416   float32      GantryTilt              # degrees

        2716   float32      XReconOffset            # mm
        2720   float32      YReconOffset            # mm

        3256   int32        BitsPerSample 
        3264   int32        DefaultWindowWidth
        3268   int32        DefaultWindowLevel GE CTI

The GE CTI family of scanners are based on the IOS platform, but fully support DICOM both on the network and on MOD media. hence it is rarely if ever desirable or necessary to get involved with the internal format within the SGI host that runs these scanners. Having said that, it is worth pointing out that internally images may be stored in a Genesis like format, with the same header layout except that some fields are 32 bit rather than 16 bit aligned (like on AW from which the IOS platform was derived), or in a true DICOM format, with a Part 10 style meta-header, except that the meta-header is encoded in implicit not explicit little endian (since it was designed and implemented before the standard Part 10 was finished and hence used the convention of early drafts).

None of this should be of consequence however, since images should always be exported from CTI scanners using network transfer or on DICOM media.

There are a few caveats however, both for the network and for media.

For network transfers, be absolutely sure that the storage SCP accepts only DICOM standard SOP classes during association negotiation, and is not promiscuous ("I will store anything of any SOP class"). Otherwise the CTI will by preference send proprietary GE SOP Classes of the ID/NET 2.0 variety, which are very DICOM like but are sufficiently different from the standard CT sop class to cause problems. The SOP Class UIDs of the ID/NET 2.0 SOP CLasses are specified in the conformance statement and if you absolutely must know what they contain there is an old service direction that describes them that is probably still available.

For the DICOM MOD media, the problems are more serious, and some of them are described in the more recent CTI conformance statement and are further explained here. Note that all these problems have been fixed, so that more recent CTI, MR LX and AW 3X devices should be writing good conformant media but still be able to read the old "bad" media". However since there may be shelves full of "bad" media one needs to be aware of the details of the problem. There is more bad CT media around than MR and AW since the fix came later to the CTI.

General details of the encapulsation and JPEG encoding are defined in DICOM Part 5 and ISO 10918-1, and explained in this FAQ in DICOM Compression. Specific details of the GE bugs are defined here, as well as being described in more recent GE CTI Conformance statements. See for example section 3.4.2 of GE Direction 2162114-100 High Speed Advantage 4.1 and 5.3 Conformance Statement.

There are two classes of problem, one related to the DICOM encapsulation, and the other to the JPEG encoding itself.

Even though all DICOM encapsulated transfer syntaxes specify little endian byte order for all non-pixel data values and for all element tags and value lengths, inadvertantly some of the delimiter and item tags in GE encapsulated pixel data are sent in either big endian for each of the group and element of the item and sequence delimited tags, or in little endian for the concatenated value of group and element as af they were a 32 bit word. That is instead of (FFFE,E000) Item being sent as FE,FF,00,EO as specified in the standard, it might be seen as FF,FE,E0,00 or 00,E0,FE,FF. Instead of (FFFE,E0DD) Sequence Delimiter being sent as FE,FF,DD,EO, it might be seen as FF,FE,E0,DD or DD,E0,FE,FF. Note also that if the Item tag is encoded wrong, then the VL field is also incorrectly encoded as a big endian 32 bit word instead of a little endian 32 bit word.

In the GE JPEG codec output, the JPEG 'SOS' header defines the Huffman table selector codes to find the appropriate Huffman table. These are incorrectly coded these as 0x11. They should have been 0x00, since those are the values assigned in the "DHI" header where the Huffman tables are actually sent. This bug manifests itself as a "Huffman table not found" error from an unpatached decoder. It also serves as a useful flag to a patched decoder that this bug (and others are present) and allows a single decoder to handle both good and bad GE compressed bit streams.

The incorrect GE JPEG computation of the difference to be Huffman encoded was computed as (Predictor - value) when it should have been calculated as (value - Predictor). The result is that the decompression with an unpatched decoder results in a "negative" of the original image. Note that GE only uses Selection Value 1 predication, so there is no need to patch other predictors.

The predictor value used at the beginning of each line used the last value of the previous line in the image, instead of the first element of the line above the current line, and for the first line, the unsigned value that is half the full scale range for the "sample precision". This manifests itself as a wierd "banding" across the image as predictions get offset by increasing errors.

An example of code that copes with both the standard and GE bugs in JPEG compression can be found in the patches to the Stanford PVRG JPEG (see JPEG Sources).

An example of code that copes with both the standard encapsluation and GE bugs in encapsulation can be found in dicom3tools "libsrc/include/pixeld/unencap.h". A section of that code (with some of the error handling removed) is reproduced here.

	size_t read(void)
			// - non-pixel data is always LE, including fragment delimiters and lengths
			// - 1st item is offset table, may have zero VL
			// - other items are fragments
			// - finally sequence delimitation tag (with zero VL)
			// - each delimiter is 2 byte group,2 byte element, 4 byte VL, little endian
			// - Item tag      is (0xfffe,0xe000) (GE mistake is 0xfeff,0x00e0 or 0xe000,0xfffe)
			// - Seq delimiter is (0xfffe,0xe0dd) (GE mistake is 0xfeff,0xdde0 or 0xe0dd,0xfffe)
			// - when GE mistake is present, fragment 32 bit VL is also swapped


			while (!lefttoreadthisfragment && !finished && !bad) {
				Uint16 group=read16();
				Uint16 element=read16();
				Uint32 vl=read32();
				if (group == 0xfffe || group == 0xfeff || group == 0xe000 || group == 0xe0dd) {
					if (group != 0xfffe) {
						cerr << "UnencapsulatePixelData::unexpected group (? bad byte order)=" << hex << group << dec << endl;
					if (element == 0xe0dd || element == 0xdde0 || group == 0xe0dd) {	// Sequence Delimiter Tag
						if (element != 0xe0dd) {
							cerr << "UnencapsulatePixelData::unexpected element (? bad byte order)=0x" << hex << element << dec << endl;
						Assert(vl == 0);
					else /* if (element == 0xe000) */ {	// Item Tag
						bool vlbyteorderwrong=false;
						if (element != 0xe000) {
							cerr << "UnencapsulatePixelData::unexpected element (? bad byte order)=0x" << hex << element << dec << endl;
						if (++fragmentnumber > 0) {
							Assert(vl);	// Zero length fragments thought not to be legal
							if (vlbyteorderwrong) {
								cerr << "UnencapsulatePixelData::assuming VL also had bad byte order, using 0x" << hex << lefttoreadthisfragment << dec << endl;
							else {
						else {
							// skip the offset table
							Assert(vl%4 == 0);
							unsigned i=0;
							while (vl) {
								Uint32 offset=read32();
				else {
					// bad tag group in encapsulated data

			if (lefttoreadthisfragment && !bad) {
				length=unsigned(lefttoreadthisfragment > maxlength ? maxlength : lefttoreadthisfragment);
				if (istr->read(buffer,length)) {
				else {

			return length;

3.2.2 Siemens CT

Some general comments about the way in which Siemens image headers, and the concept of native file formats and exported SPI formats are to be found in the section on Siemens MR. Siemens Somatom DR

This description pertains to the DR family, and possibly also earlier Siemens CT models, but I have no files from these to test.

The files are in fixed format (cf. the early Magnetom format which is similar, but has block pointers) with three major blocks of entries:

        - binary data      - offset 0    - 512 bytes
        - text overlay     - offset 512  - 960 bytes plus 676 bytes free
        - image pixel data - offset 2048 - 131072 bytes

The binary data block is filled with the usual cryptic enumerated values and useful parameters. Some of the more interesting ones are:

        - binary data block:

                66  - byte        - archive mode (0=raw data,B=256,C=512)
                67  - byte        - archive mode (0=uncompressed,

                72  - short       - matrix size (256 or 512)

                130 - byte        - scan mode (P=image data,R=raw data)
                131 - byte        - scan mode (0=tomogram,Q=quick,S=serial,
                132 - short       - fov - mm
                134 - short       - scan time - secs * 10
                136 - short       - kv
                138 - short       - dose - maS
                140 - short       - slice thickness - mm
                142 - short       - gantry tilt - degrees
                144 - short       - table position - mm
                146 - short       - table height - mm
                148 - short       - scan mode (1=standard(360),

                236 - short       - view direction (1=cranial,-1=caudal)
                238 - byte        - head position (0=head first,
                                    1=feet first)
                239 - byte        - patient position (0=supine,
                                    1=prone,2=r lat dec,3=l lat dec)

                310 - short       - window width  A
                312 - short       - window center A
                314 - short       - window width  B
                316 - short       - window center B

Unfortunately, the patient identification information is NOT stored in the binary data block, rather one has to extract it from the image text overlay block, which consists of 960 characters (24 lines of 40 characters WITHOUT carriage control characters) in a fixed format. This is where what you see overlayed on the filmed images is stored. Some of these values are duplicates of what is in the binary data block, but things like the patient name and so on are here and nowhere else :(


                0   SOMATOM DR2       ST. ELSEWHERE GEN HOSP
                40  999999-9999  JOHN DOE                EF2
                80  01-JAN-90        FRONT               35B
                120 13:31:22                            H/SP
                200 SCAN 60                             L   
                240                                     E   
                280                                     F   
                320                                     T   
                720 TI 5                                    
                760 KV 125                                  
                800 AS .35                                  
                840 SL 2                                    
                880 GT 0                                    
                920 TP 144                                  

        - text overlay block: (some of this is guess work)

                0   - char[14]    - product
                15  - char[25]    - hospital name
                40  - char[12]    - patient number
                53  - char[22]    - patient name
                80  - char[2]     - date - dd
                83  - char[3]     - date - mmm
                87  - char[2]     - date - yy
                120 - char[2]     - time - hh
                123 - char[2]     - time - mm
                126 - char[2]     - time - ss
                156 - char[1]     - H=head first,F=feet first
                158 - char[2]     - SP=supine,PR=prone,
                                    RP=right lateral decubitus,
                                    LP=left lateral decubitus
                205 - char[4]     - slice number
                723 - char[4]     - scan time - secs
                763 - char[4]     - kv
                803 - char[4]     - dose - AmpS
                843 - char[4]     - slice thickness - mm
                883 - char[4]     - gantry tilt - degrees
                923 - char[4]     - table position - mm

If anyone knows what "EF2" and "35B" stand for I would love to know - I presume they are something like the filter used, or field of view or something ?

Also the DR family don't seem to be aware of the concept of a hierarchy of examination/study and series numbering, which makes it annoying to try to import them into PACS systems :( Correct me if I am wrong but they just seem to keep bumping up the slice number for each patient as each group of scans is done. Siemens Somatom Plus

There seem to be different formats for different versions of the machine. Either that or some sites have PACS software and some don't or something. Anyway, one set of files that were sent to me used a fixed format header much like the DR family, but of different length and with different fields. I have not yet adequately deciphered this header but will include it here when I have. This may be what is referred to as the "original header" stored in the SPI format.

Another site uses a Siemens version of SPI, containing the following private data elements. Note that there is overlayed data in the high four bytes of the image pixel data, and that there seems to be a bunch of padding in the middle. The intent seems to be to store the "original header" and the image pixel data at accessible, presumably standard locations, presumably indexed by the byte offsets and lengths described in group 9. This is a shame because it seems that none of the really interesting CT attributes have been included in the SPI form, although SPI private tags are available for lots of CT parameters. I don't have one of these image to test this theory, someone just sent me an output of the attribute dump.

SPI private tags:

(0009,0010)                                      <SPI RELEASE 1>
(0009,0011)                                        <SIEMENS MED>
(0009,1011) SPI RELEASE 1   UID     <049S03CT031995011712072452>
(0009,1040) SPI RELEASE 1   DataObjectSubtype           [0x0000]
(0009,1041) SPI RELEASE 1   DataObjectSubtype         <IMA TOPO>
(0009,1110) SIEMENS MED     RecognitionCode             <CT 1.4>
(0009,1130) SIEMENS MED     ByteOffsetOfOriginalHeader
(0009,1131) SIEMENS MED     LengthOfOriginalHeader
(0009,1140) SIEMENS MED     ByteOffsetOfPixelmatrix
(0009,1141) SIEMENS MED     LengthOfPixelmatrixInBytes 

(0011,0010)                                      <SPI RELEASE 1>

(0021,0010)                                        <SIEMENS MED>
(0021,1010) SIEMENS MED     Zoom                          <01.0>
(0021,1011) SIEMENS MED     Target              <000.000\00.000>
(0021,1012) SIEMENS MED     TubeAngle                     <0270>
(0021,1020) SIEMENS MED     ROIMask                     [0xf000]

Overlay descriptions (overlays already in image pixel data):

(6000,0040)                 ROI                              <G>
(6000,0102)                 BitPosition                 [0x000c]
(6000,0102)                 OverlayLocation             [0x7fe0]

(6002,0040)                 ROI                              <G>
(6002,0102)                 BitPosition                 [0x000d]
(6002,0102)                 OverlayLocation             [0x7fe0]

(6004,0040)                 ROI                              <G>
(6004,0102)                 BitPosition                 [0x000e]
(6004,0102)                 OverlayLocation             [0x7fe0]

(6006,0040)                 ROI                              <G>
(6006,0102)                 BitPosition                 [0x000f]
(6006,0102)                 OverlayLocation             [0x7fe0]

More SPI private stuff ... padding and original header ...

(7001,0010)                                        <SIEMENS MED>
(7001,1010) SIEMENS MED     Dummy

(7003,0010)                                        <SIEMENS MED>
(7003,1010) SIEMENS MED     Header

(7005,0010)                                        <SIEMENS MED>
(7005,1010) SIEMENS MED     Dummy Siemens Somatom AR


3.2.3 Philips CT - Big black hole

3.2.4 Picker CT

Grey hole perhaps. This information probably pertains to the IQ and PQ CT models, though I have no sample images to experiment with yet. I am told that:

3.2.5 Toshiba CT - another black hole

3.2.6 Hitachi CT - another black hole

3.2.7 Shimadzu CT - another black hole

3.2.8 Elscint CT - another black hole

3.2.9 Imatron CT

The following information is included verbatim from that kindly supplied by Cameron Ritchie:

Imatron File Format

In this document, the Imatron file format is described. Imatron makes no guarantees that future Imatron files will be compatible with the attached format. This format is current as of 2/29/96.

The format described here is generally true for files produced by all Imatron scanners (C-100, C-150L, C-150, C-150XP, C-150LXP); however, some small differences may be found. The file format described below is valid for image files on the scanner's RT-11 disks. What is not described is how to actually get one of these files off the RT-11 and on to a workstation or PC for conversion. This procedure is actually almost more difficult than the conversion! There are three options for getting files off the scanner; only one does not require additional hardware. The options are as follows:

Two demo image extractors are available for download. Both are available with source code, and Imatron does not guarantee either program's accuracy. The first program converts Imatron format files to headerless files. The second program converts Imatron files to Siemens Somatom, headerless, DICOM, or TIFF. Command line help can be obtained for either program by typing program_name -h.

Imatron hopes that the information contained here is useful to the research community. Assistance, within reason, can be obtained by contacting:

Cameron J. Ritchie, Ph.D.
Applications Scientist
Imatron Inc.
389 Oyster Point Blvd.
South San Francisco, CA 94080

Disk Data-File Formats

Scan data collected are stored as raw data in files on the VME disk drive. After reconstruction they are stored as image data files on the RT-11 disks. These files comprise header information and the acquired data. An Imatron file is a set of information about multiple slices. Each file contains:

Block 0: Control Block

The control block is the first block of an Imatron file and contains information necessary for interpreting the rest of the file (Table 2-1).

Table 2-1: Words in a Control Block

   WORD                         DESCRIPTION                      

     0      Pointer to first block in the file header            

     1      Number of entries in the file header                 

     2      Pointer to first block of the file header data       

     3      Pointer to first block in the slice header           

     4      Number of entries in the slice header                

     5      Pointer to first block of the slice header position  

     6      Number of words in a header table entry              

     7      File type version number                             

     8      Number of blocks of detector offset data             

     9      Number of blocks in file header table                

    10      Number of blocks of file header data                 

    11      Number of blocks in slice header                     

    12      Number of blocks in slice header position table      

    13      Number of blocks for each section of slice header    

    14      Pointer to start of detector offset blocks           

  15-255    0                                                    

File and Slice Headers

Imatron file and slice headers store information about: file organization, the patient, scanning, reconstruction, and how to perform image analysis on the data. Information in these headers is not stored in fixed locations in the file. Instead, there is a symbol table that references the header values by name. There are two symbol tables in each file: the file-header symbol table (referred to as the file header or file-header table), containing names and pointers into a single file-header data area; and the slice-header symbol table (referred to as the slice header or slice-header table), which uses the same format but its pointers are used for all the slice-header data areas (one per slice).

The file header and the slice header are composed of pointer/descriptor units which point to variables in the data blocks. Each unit is 6 words (12 bytes) long and organized as shown in Table 2-2.

Table 2-2: Unit Organization

   BYTE                           Contents                         

   1-6     ASCII variable name, padded with null bytes             

    7      Null byte (0)                                           

    8      ASCII variable type (I => Integer, B => Byte, F =>      
           Floating Pt.)                                           

   9-10    Integer pointer to the word number in the block where   
           the data for this variable starts                       

  11-12    Number of data values of the type described in byte 8.  

The integers contained in bytes 9-10 and 11-12 are stored with the least significant byte in the first byte, and the most significant byte in the second byte.

The following is an example of how the file-header parameter ICMNTS is defined:

BYTE:     1-6         7       8     9      10     11     12     

          ICMNTS      0      'B'    37     0      80     0      

Parameter variables:

One block can contain up to 42 pointer/descriptor units.

Slice-Header Position Table

The slice-header position table contains a list of unsigned integer pointers to the various slice-header data blocks. The first word of this table points to slice-header data block 1, the second to slice-header data block 2, etc.

ECG Data

ECG data is stored in the raw (.VME) and image files for ECG-triggered studies. The file header variable ITRTYP, points to the starting block in the file for this set of data, which, if present, is 32 blocks long. There is no slice header associated with the data.

File header parameters are shown in Table 2-3; slice headers are shown in Table 2-4.

Image Data-File Formats

The C-150 scanner produces axial slices by sweeping an electron beam along one of four target rings (Target A, B, C, or D). X-rays produced by the scanning electron beam are detected by a pair of solid-state detector rings (Detector Rings 1 and 2).

In an N-image (Imatron image) file there are N slices, 1 slice per image. The slice-header parameters, NROWS and NCOLS, define the number of rows and columns in the stored rectangular image. Data is not compressed. The first NCOLS words in the slice are the first row, the second NCOLS words are the second row, etc. Image data are converted to Hounsfield units by subtracting 1000 (decimal) from each word. The resulting numbers range from -1000 to +3095 inclusive (Imagraph).

Table 2-3: Data File Header Format

 INDEX  NWDS      NAME                  DESCRIPTION              

   1      1      IFHLEN       The number of 256 word blocks in the  
                              file header.                          

   2      1      ISHLEN       The number of 256 word blocks in the  
                              slice header.                         

   3      5      IAFN         The ASCII file descriptor.  (6 char.  
                              name,'.',3 char. extension)           

   8      5      IADATE       ASCII date string. (9 character       
                              string right-padded with a blank.     

  13      4      IATIME       ASCII time string. (8 character       

  17      6      IPATID       ASCII patient ID number. (12 chars.)  

  23     15      IPATNA       ASCII patient name. (30 chars.)       

  38     40      ICMNTS       ASCII comments. (80 chars.)           

  78      1      NDETS        The number of detectors. (432 or      

  79     63      IDEMAP       The detector status map for the       
                              file.  All bits defined as            
                              1=working, 0=inoperative. Channel     
                              k's status is indicated in word       
                              IW=1+(k-1)/16, [integer arith.] of    
                              IDEMAP, by bit                        
                               IBIT = k - (IW-1)*16 - 1.            

  142     1      ISTOB        The starting block for detector       
                              offset measurements. (0 for no        
                              offsets recorded.)                    

  143     1      NSLICE       The number of slices in the file.     

  144     1      IORGAN       The file organization code:           

                              -2 = unsorted raw MM data (AIR, PIN   
                              or OFFSET)                            
                              -1 = unsorted raw MM data             
                              0 = source-fan data                   
                              1 = detector-fan data                 
                              2 = image (rectangular) data          
                              3 = tuning point data                 
                              4 = deflection buffer data            
                              5 = processed calibration data        
                              6 = processed AIR data                
                              7 = processed OFFSET data             

145       1      ITTICK       The DAS clock period is               

146       1      NPHVEW       The number of phantoms.               

147       1      IDATYP       0 = DAS output words (All RAW data)   
                              1 = Integer                           
                              2 = Floating point (Sinogram,         
                              3 = Scaled 11-bit integer data        
                              (image & screen save)                 
                              4 = AP400 block floating point        
                              5 = MM address data (calibration      
                              6 = Octal data (deflection buffer     
                              7 = Packed Fast Raw Averaged Data     
                              8 = Scaled 12-bit integer data        
                              (image & screen save)                 

148       1      NDETOM       No. of detector offset measurements   

149       2      XMMTMU       The scale factor to change from mm    
                              to MIP machine units (units are       

151       1      IREP         The no. of DAS samples per detector   
                              per source fan. IREP = 3 for a 50ms   
                              scan, IREP = 6 for a 100ms scan.      

152       2      PIXLEN       Length in mm. of a pixel, from        

154       1      NLEVEL       Number of levels in the file          

155       1      NPLEVL       Number of images per level (valid in  
                              raw image files. Level number is an   
                              integer from 1 to NLEVEL.  Closest    
                              to the gun is first.)                 

156       1      IREF         2 Byte ASCII description of the       
                              reference pt.                         

157       1      ISTUDY       Study type:                           

                              -199 to -100 reserved for test &      
                              calibration "studys" (Not             
                              -1 = SCREEN SAVE   => No analysis     
                              0 = SPECIAL STUDY => Anything not     
                              below.  Atypical study                
                              1 = LOCALIZATION  => single scans, 2  
                              images per scan, N scans at           
                              arbitrary levels (in pairs) (50 ms)   
                              2 = FLOW STUDY    => Typically, a     
                              set of scans triggered periodically.  
                              3 = MOVIE STUDY   => Typically, many  
                              scans taken continuously.             
                              4 = AVERAGE VOLUME=> Averaged data    
                              from a volume study on a single       
                              target ring                           
                              5 = VOLUME STUDY  => various times    
                              at lots of levels (table motion)      
                              6 = AVERAGE FLOW  => Averaged data    
                              from a flow study on a single target  
                              7 = CONTINUOUS VOLUME (CVS)           

                              51 = IMAGE AVERAGING                  
                              52 = REFORMAT                         
                              53 to 61 reserved for FUNCTIONAL      
                              IMAGE PROC.                           
                              53 = FIP Maximum Difference           
                              54 = FIP Time to Peak                 
                              55 = FIP Area Under the Curve         
                              56 = FIP Center of Mass               
                              102 = IMAGE SUBTRACTION FLOW          
                              103 = IMAGE SUBTRACTION MOVIE         
                              105 = IMAGE SUBTRACTION VOLUME        
                              106 = IMAGE SUBTRACTION AVERAGE FLOW  

158      10      ICONTR       Type of contrast (20 characters)      

168       2      DOSECN       Contrast dose in cc                   

170      10      INJSIT       Injection Site (20 characters)        

180      10      ISTRES       Type of stress (20 characters)        

190       7      IRPHYS       Referring physician's last name (14   

197       7      IRADIO       Radiologist's last name (14 chars)    

204       2      ITECH        Radiation technologist's initials (3  

206       5      IBDATE       Patient's birthdate (9 chars (ex.     

211       1      ISTHCK       Slice thickness, mm.                  

212       1      ICALIB       Calibration number                    

213       1      KERNEL       Desired kernel flag                   

214       1      ITRTYP       trigger type:                         
                              1 = manual                            
                              2 = timed                             
                              3 = ecg with no extra data else it's  
                              a pointer to ecg data in the file     

215       1       IPATSZ      patient size:                         
                              1 = small                             
                              2 = medium                            
                              3 = large                             
                              4 = shoulder/pelvis kluge             

216       1      IPRLVL       regular reconstruction's first level  
                              to recon:0 = none  else 1 to nlevel   

217      20      IDIAG        diagnosis comment                     

237       9      IHOSP        hospital (actually scanner)           

246       4      BOLTIM       Bolus times                           

250       1      NSPLIT       Number of images to be created from   
                              each raw slice                        

251       1      IDLINP       Delete raw data flag:                 
                              0 = do NOT delete after recon         
                              1 = delete after complete recon       

252       2      CDENS        Density of contrast                   

254       1      IOFMIN       Time since midnight in minutes of     
                              last offsets                          

255       1      IOFDAT       Day since dec 31 of last offsets      

256       1      NRINGS       Number of detector rings used.        

257       1      NTARGT       Number of targets used.               

258       1      ICNREC       0 = not suitable for cone beam        
                              1 = suitable for cone beam            
                              2 = suitable and cone beam alg used.  

259       6      KERNAM       ASCII kernel name used.               

260       1      ISNTYP       Sinogram type.                        

261       1      IANTYP       Analysis type for ASA                 
                              1 = Cone analysis                     
                              2 = Air analysis                      
                              3 = Pin analysis                      

262       1      ISTHCF       Slice thickness.  LSB = 1/100 mm.     

263       1      ICOLL        Collimator setting (1=1.5mm, 3, 6)    

Table 2-4: Data Slice Header Format

 INDEX     NWDS     NAME                  DESCRIPTION              

   1        1      ISDATP      Pointer to data for this slice       
                               (Always here!)                       

   2        2      R1MU        Linear attenuation co-efficient for  
                               water at this energy and current,    
                               ring 1.                              

   4        1      IROTA       = 1 clockwise scan,                  
                               = -1 for counter-clockwise scan.     

   5        2      HVDES       Desired high voltage for this scan,  
                               in kV.                               

   7        2      HVACT       Actual high voltage for this scan,   
                               in kV                                

   9        1      ICURNT      Actual electron beam current, in     

  10        2      FVDES       Desired filament voltage, in volts.  

  12        2      FVACT       Actual filament voltage, in volts.   

  14        2      FCACT       Actual filament current, in          

  16        1      IRING       The detector ring used:              
                               0 = Raw slice with both RINGs        
                               1 = RING 1 (closest to gun)          
                               2 = RING 2 (farther from gun)        

  17        1      ITARGT      The target ring used.                

  18        1      NSLAVG      The number of scans averaged to      
                               produce this slice.                  

  19        2      PICRAD      Floating point picture radius in     

  21        2      XORG        Floating point X coordinate of       
                               reconstruction center (0.0 is        
                               isocenter) in mm.                    

  23        2      YORG        Floating point Y coordinate of       
                               reconstruction center (0.0 is        
                               isocenter) in mm.                    

  25        2      ZOOM        Floating point zoom factor (1.0 =    
                               no zoom) for reconstruction          

  27        1      NROWS       The number of rows in the            
                               reconstructed image.                 

  28        1      NCOLS       The number of cols in the            
                               reconstructed image.                 

  29        2      VALMAX      Maximum value in the slice (in       
                               floating point)                      

  31        2      VALMIN      Minimum value in the slice (in       
                               floating point)                      

  33        2      RSCALE      Data has been scaled and biased      
                               such that                            

  35        2      RMIN        actual data = data/RSCALE + RMIN     

  37        1      IPATH       Holding path flag:                   
                               0 = path was HOLDING PATH            
                               1 = path was the first for that      
                               2 = the slice was NOT the first of   
                               that pulse (slices 2-N for a movie   
                               or volume)                           

  38        2      ELAPSE      Time, in seconds, since the first    

  40        1      LEVELN      The level number for a given slice   

  41        2      ISTAGE      Old:2 word array, 2nd word unused,   
                               1st word is >=0 if data is present   
                               and useful.                          

  43        1      INOUT       In-out table pos. relative to ref.   
                               (-0.1 mm)                            

  44        1      IHITE       Up-down table pos. relative to       
                               reference (mm)                       

  45        1      ITILT       Table tilt relative to horizontal    

  46        1      ISLEW       Table slew relative to straight      

  47        1      ICPHAS      Cardiac phase in % R-R-wave          

  48        1      IBEAT       Heart beat # for this image          

  49        2      HRATE       Heart rate in beats per minute       

  51        1      IPATOR      Integer code for patient             
                               0 = not applicable or special case   
                               5 = prone head first flipped         
                               + 1 = supine                         

                                + 2 = prone                         
                                + 3 = decubitus right               
                                + 4 = decubitus left                
                                -5 = supine ff (flipped to match    
                                -6 = prone ff (ditto 2)             
                                -7 = decub right (ditto 3)          
                                -8 = decub left (ditto 4)           
                               Positive refers to HEAD FIRST (head  
                               closest to gun).  Negative refers    
                               to FEET FIRST.                       

  52        2      SLSIZE      Size of slice in words               

  54        1      ITN         Order of Chebychev polynomial        
                               applied to data (only if valid       
                               during calibration, for normal       
                               recon ITN = 0).                      

  55        2      R2MU        Linear attenuation coefficient for   
                               water at this energy and current,    
                               ring 2.                              

  57        1      IVMFLAG     Contains bit-map of flags used by    

  58        1      NTARGS      Number of target sections of this    
                               target ring.                         

Scanner Operating Modes

The scanner operates in two different modes: Single-Slice Mode (SSM) and Multi-Slice Mode (MSM).

Single-Slice Mode:

The FILE HEADER variable "IREP" defines Single-Slice Mode:

IREP = 6 for SSM

The total number of images in the file is the FILE HEADER variable NSLICE.

The total number of axial slice positions in the file is the FILE HEADER variable NLEVEL.

In SSM, only Target Ring C and Detector Ring 2 are used.

Each sweep of the beam along Target Ring C takes 100 milliseconds.

The exposure time (in seconds) is determined by the SLICE HEADER variable "NSLAVG":

Exposure time (seconds) = NSLAVG * 0.1

The axial position for each slice is determined by the SLICE HEADER variable "INOUT" (which is in tenth mm units):

Slice position relative to reference (in mm) = INOUT/10

Multi-Slice Mode:

The FILE HEADER variable "IREP" defines Multi-Slice Mode:

IREP = 3 for MSM

The total number of images in the file is the FILE HEADER variable NSLICE.

The total number of axial slice positions in the file is the FILE HEADER variable NLEVEL.

In MSM Mode, each sweep of the electron beam along a single target ring produces a pair of simultaneously acquired, side-by-side axial slices (1 from each detector ring).

Any combination of target rings (A, B, C, or D) may be used.

Each sweep of the beam along any single target ring takes 50 milliseconds.

The exposure time (in seconds) is determined by the SLICE HEADER variable "NSLAVG":

Exposure time (seconds) = NSLAVG * 0.05

The axial position for each slice is determined by the SLICE HEADER variables "INOUT," "ITARGET," and "IRING" and may be calculated as follows:

KTARGT = ITARGT - 64 /* Convert ascii target to integer */

TAROFF = -20.0 + (4 - KTARGT)*20.0 /* Distance from C to target */

DETOFF = mod(IRING,2)*8 /* Distance from detector Ring 2 to detector */

Slice position relative to reference (in mm) = INOUT/10. + TAROFF + DETOFF

Study Types

The six Imatron study types are described as follows:

SSM Flow (IREP = 6, ISTUDY = 6)

Description:         For an N-slice SSM Flow Study, the following  
                     is repeated NSLICE times:  A 100-ms sweep of  
                     the beam is performed NSLAVG times along      
                     ring C (with 16 ms between sweeps), and the   
                     data for the NSLAVG sweeps are summed         
                     together to produce a single image.  All of   
                     the data are acquired at a single axial       
                     slice position, sequentially in time.         

File Organization:   Slice 1 in the file is the first "time,"      
                     slice 2 is the second "time," ...slice n is   
                     the nth "time."                               

SSM Cine (IREP = 6, ISTUDY = 3)

Description:         For an N-slice SSM Cine study, NSLICE 100-ms  
                     sweeps of the beam are performed along        
                     Target Ring C (with 16 ms between sweeps).    
                     Each sweep of the beam produces a single      
                     image.  All of the data are acquired at a     
                     single axial slice position, sequentially in  

File Organization:   Slice 1 in the file is the first "time,"      
                     slice 2 is the second "time,"  ... slice n    
                     is the nth "time."                            

SSM Volume (IREP = 6, ISTUDY = 5)

Description:         In an N-slice volume study, the following     
                     sequence is repeated NSLICE times in          
                     succession:  A 100 ms sweep of the beam is    
                     performed NSLAVG times along ring C (with     
                     16-ms between sweeps), and the data for the   
                     NSLAVG sweeps are summed together to produce  
                     a single image.  Then the patient table       
                     moves to a new axial position.                

File Organization:   Slice 1 in the file is the first "level,"     
                     slice 2 is the second "level," ... slice n    
                     is the nth "level."                           

MSM Volume (IREP = 3, ISTUDY = 5)

Description:         MSM Volume Studies always use Target Ring C   
                     (only), and both Detector Rings 1 and 2.  In  
                     an MSM VOLUME STUDY, the following sequence   
                     is repeated NLEVEL/2 times in succession:  A  
                     50-ms sweep of the beam is performed NSLAVG   
                     times along ring C (with 8-ms between         
                     sweeps) and the data for the NSLAVG sweeps    
                     are summed together to produce a pair of      
                     side-by-side images acquired at adjacent      
                     axial positions.  Then the patient table      
                     moves to a new axial position.                

File Organization:   Slice 1 in the file is the first "level,"     
                     slice 2 is the second "level," ...slice n is  
                     the nth "level."                              

MSM Flow (IREP = 3, ISTUDY = 1 or 2)

Description:         Refer to the general Multi-Slice Mode         
                     description, above. In an MSM Flow study,     
                     the following applies:                        

                     N_T = The number of times = NSLICE/NLEVEL     

                     The following action is repeated N_T times:   

                       For a 2-level MSM Flow, the beam sweeps     
                     once on a single target ring (A, B, C, or D)  
                     to produce a pair of side-by-side images      
                     acquired at the same "time."                  

                       For a 4-level MSM Flow, the beam sweeps     
                     once on one target ring, then 8 ms later,     
                     sweeps on a second target ring; this          
                     produces 4 side-by-side images acquired at    
                     the same "time."                              

                       For a 6-level MSM Flow, the beam sweeps     
                     once on one target ring, then 8 ms later,     
                     sweeps on a second target ring; followed 8    
                     ms later by another sweep, on a third target  
                     ring; this produces 6, side-by-side images    
                     acquired at the same "time."                  

                     For an 8-level MSM Flow, the beam sweeps      
                     once on one target ring, then 8 ms later,     
                     sweeps on a second target ring; followed 8    
                     ms later by another sweep on a third target   
                     ring, and again, 8 ms later on the fourth     
                     target ring; this produces 8, side-by-side    
                     images acquired at the same "time" (Table     

Table 2-5: File Organization for MSM Flow

  Slice in File     Time            Axial Position            

        1             1     use MSM Template info => axial       
                            index i                              

        2             1     use MSM Template info => axial       
                            index ii                             

        3             1     use MSM Templace info => axial       
                            index iii                            

        .             .     .                                    

        .             .     .                                    

        .             .     .                                    

     NLEVEL           1     use MSM Template info => index       

    NLEVEL+1          2     axial index i                        

    NLEVEL+2          2     axial index ii                       

        .             .     .                                    

        .             .     .                                    

        .             .     .                                    

    2*NLEVEL          2     axial index nlevel                   

   2*NLEVEL+1         3     axial index i                        

        .             .     .                                    

        .             .     .                                    

        .             .     .                                    

(N_T-1)*NLEVEL+1     N_T    axial index i                        

(N_T-1)*NLEVEL+2     N_T    axial index ii                       

        .             .     .                                    

        .             .     .                                    

        .             .     .                                    

   N_T*NLEVEL        N_T    axial index nlevel                   

MSM Cine (IREP = 3, ISTUDY = 3)

Description          Refer to the general MSM description above.   
                     In an MSM Cine study, the following applies:  

                     N_T = The number of times = NSLICE/NLEVEL     

                     NTARGS = The number of targets used =         

The following action is repeated NTARGS times (once for each target): N_T 50-ms sweeps of the beam are performed, with 8-ms between sweeps, along a single target ring (A, B, C, or D); this produces a pair of images acquired at adjacent axial positions. (See Table 9, below.)

Table 2-6: File Organization for MSM Cine

Slice in File       Time              Axial Position             

      1               1     use MSM Template info => axial index  

      2               1     use MSM Template info => axial index  

      3               2     axial index i                         

      4               2     axial index ii                        

      5               3     axial index i                         

      6               3     axial index ii                        

      .               .     .                                     

      .               .     .                                     

   2*N_T-1           N_T    axial index i                         

   2*N_T             N_T    axial index ii                        

   2*N_T+1            1     use MSM Template info => axial index  

   2*N_T+2            1     use MSM Template info => axial index  

   2*N_T+3            2     axial index iii                       

   2*N_T+4            2     axial index iv                        

      .               .     .                                     

      .               .     .                                     

   4*N_T-1           N_T    axial index iii                       

   4*N_T             N_T    axial index iv                        

      .               .     .                                     

      .               .     .                                     

(NLEVEL-2)*N_T+1      1     use MSM Template info => axial index  

(NLEVEL-2)*N_T+2      1     use MSM Template info => axial index  

(NLEVEL-2)*N_T+3      2     axial index nlevel-1                  

(NLEVEL-2)*N_T+4      2     axial index nlevel                    

      .               .     .                                     

      .               .     .                                     

NLEVEL*N_T-1         N_T    axial index nlevel-1                  

NLEVEL*N_T           N_T    axial index nlevel                    

The next part is part4 - proprietary MR formats.