This is fascinating and seems to explain the confusion. Thanks for that. What is informed wisdom's advise for compressing CTs for remote reporting? I assume that one shouldn't compress plain films but am slightly in the dark!! Tom Naunton Morgan
posted on Wednesday, January 17, 2007 - 12:57 pm
In general, clinical studies so far show that you CAN compress plain films and CT using moderate LOSSY compression without losing any significant clinical information. There is an on-going Canadian study to clarify where the cut-off exists, and to test different algorithms, but no studies so far have shown any loss of diagnostic accuracy at about 10:1 for low resolution images (CT, MR etc.) and at 20:1 for high matrix images such as plain films, using standard JPEG (12 bit where necessary). So, contrary to what you might think, plain films can sensibly be compressed MORE than CT/MR etc.
Not "treading on my toes" at all - I don't "own" this subject! .....at least we gave the same answer!
Dear Dave Which are the DICOM recognised forms of lossy compression? Where do the various wavelet compression algorithms stand in all this please? Thank you Nicola
posted on Wednesday, January 17, 2007 - 02:43 pm
OK - here's the full list - including lossless. I include the DICOM transfewr syntax UIDs for completeness, but most people probably don't need to know them!
JPEG Lossy ========== Moderately well supported in practice. 2 forms:
"Baseline": Only allows 8 bit data : 1.2.840.10008.1.2.4.50 "Extended": Allows 12 bit data : 1.2.840.10008.1.2.4.51 This second version allows use for rewindowable images - CT is only 12 bits anyway, and MR can be "rescaled" to fit if necessary
JPEG Lossless ============ Very good industry support, and used internally by many systems. There are 2 forms, one having slightly restricted options to make life easier for simple implementations, but therefore with slightly less compression. That is the one widely used in cardiology systems.
Run-Length Encoding (RLE) ========================= Simplistic lossless method only suitable where there are large areas of black (e.g. suitably encoded U/S). Rarely used 1.2.840.10008.1.2.5
JPEG 2000 ========= Only slowly coming into use. Unlike older JPEG, this uses the SAME basic mechanism for both lossy & lossless, the only difference being that lossy is a "truncated" version of the lossless data stream. This IS a wavelet algorithm, and is the only standardised interoperable wavelet method. Despite the hype, there is little difference in practice between usable compression ratios for original JPEG and JPEG 2000, and some concerns have been raised that artefacts in JPEG 2000 look "natural" (and are therefore dangerous) as opposed to JPEG, where artefacts are "blocky" and obvious. The major potiential benefit of JPEG 2000 is for streaming applications, but few vendors have yet to implement non-proprietary streaming using it. Two different transfer syantxes to allow negotiation: Lossless : 1.2.8126.96.36.199.90 Lossless OR lossy : 1.2.840.10008.1.2.4.91
JPEG-LS ======= Added a few years back, with a few influential backers, but never really used: 1.2.840.10008.1.2.4.80 1.2.840.10008.1.2.4.81
MPEG ==== This is the same as used for DVDs, digital TV etc., and is for video, including endoscopy etc. It gives MUCH higher compression ratios for the same quality, as it uses the frame-to-frame similarities in the data, unlike the other methods, which work "one frame at a time". Slowly coming into use: 1.2.840.10008.1.2.4.100
There have also been multiple "Private" transfer syntaxes/compression methods over the years, offering wavelet before JPEG 2000 became available. Though salemen remain stubbornly keen on claiming that THEIR wavelet mechanism is however SPECIAL, there is little if any evidence to suggest that any proprietary versions are any better than JPEG 2000, and they certainly wouldn't have enough benefit to offset the loss of standardised interoperability. They are already falling out of use, being superceded by JPEG 2000, and this will continue.
Is this perhaps too much information?
posted on Wednesday, January 17, 2007 - 03:48 pm
The less 'techy' amongst us may like to see the SCAR review of the excellent Canadian work on compression for radiology images. They promise more information (also see a previous posting in this topic) but for CT it looks as if (from this work and others) that 5:1 JPEG2000 and for CR 20:1 JPEG200 are OK. There is an interesting legal caveat in the SCAR report and as Canadian law is similar to UK law this interpretation should be heeded!
posted on Wednesday, January 17, 2007 - 03:52 pm
Comprehensive information, nicely presented.
One small typo to correct - CT is only 8 bits, rather than 12.
posted on Wednesday, January 17, 2007 - 04:13 pm
Doh! Apology needed I think. Spoke too soon - 8 bits made Dave's statement make more sense, but 12 bits (as he said) is correct.
In fact, with the new CT scanners using extended CT number range, it can be even more than 12 bit.
posted on Wednesday, January 17, 2007 - 04:19 pm
NO!!! - CT is at least 12 bits !!! (and if more can be rescaled like MR)
[Reference: DICOM PS3.3 C.188.8.131.52.5 : "For CT Images, Bits Stored (0028,0101) shall have the Enumerated Values of 12 to 16."]
Where does your 5:1 figure come from? (which is barely more than the 3:1 available lossless!) The Canadian SCAR article (and all others I have previously seen) have found 10:1 OK for CT.
posted on Wednesday, January 17, 2007 - 04:27 pm
Apology accepted (I wrote the last message before your 2nd comment appeared). To clarify my original comment (which was perhaps not well explained):
The 12 bit (extended) JPEG mode allows enough bit depth to be preserved for meaningful re-windowing, which is not true for the 8 bit mode. Most CT data is 12 bits, and in fact most MR (though nominally 16 bits) also fits into a 12 bit range. For data that does not, they can and should be rescaled prior to compression - whilst this does potentially lose a few "least significant bits", in the context of lossy compression anyway, this is unlikely to be significant.
Most of the other compression types listed support any number of bits (including JPEG 2000 lossy), the only exception being MPEG which is limited to 8 bits per colour.
posted on Wednesday, January 17, 2007 - 04:34 pm
Can't remember at present, but I think it was from a recent ER paper on chest CT. I'll do a bit of browsing and see if I can find the ref.
posted on Wednesday, January 17, 2007 - 04:43 pm
posted on Wednesday, January 17, 2007 - 05:12 pm
That article is looking at a slightly different question from that normally asked.....it is looking at whether the images are "distinguishable" as opposed to the more clinically relevant measure of "equal diagnostic use", which is what more sophisticated studies such as the Canadians are using. As an example of the difference, it is often commented (as in the SIIM article) that moderately compressed images have a little less noise, and are sometimes "Preferred". Such images would be rated as distinguishable (and fail the test on this paper) but pass diagnostically based tests.
posted on Wednesday, January 17, 2007 - 05:55 pm
Aye, well there you go. I guess we (RCR IT Committee?)are going to have to determine which way to fall off the fence! Whichever way it is going to have to be legally defensible and defensible to radiologists who might just perceive a difference and be unsettled? The final outcome of the excellent Canadian work may determine this.