Pergamon
`
`CMIG 264
`
`Computerized
`MedicalImaging
`andGraphics
`
`Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`A comparison of lossless compression methods for medical images
`Juha Kivija¨rvia, Tiina Ojalaa, Timo Kaukorantaa, Attila Kubab, La´szlo´ Nyu´lb, Olli Nevalainen a,*
`aTurku Centre for Computer Science (TUCS), Department of Computer Science, University of Turku, Lemminka¨isenkatu 14 A, FIN-20520 Turku, Finland
`bDepartment of Applied Informatics, Jo´zsef Attila University, Szeged, PO Box 652, H-6701 Szeged, Hungary
`
`Abstract
`
`In this work, lossless grayscale image compression methods are compared on a medical image database. The database contains 10 different
`types of images with bit rates varying from 8 to 16 bits per pixel. The total number of test images was about 3000, originating from 125
`different patient studies. Methods used for compressing the images include seven methods designed for grayscale images and 18 ordinary
`general-purpose compression programs. Furthermore, four compressed image file formats were used. The results show that the compression
`ratios strongly depend on the type of the image. The best methods turned out to be TMW, CALIC and JPEG-LS. The analysis step in TMW is
`very time-consuming. CALIC gives high compression ratios in a reasonable time, whereas JPEG-LS is nearly as effective and very fast.
`q 1998 Elsevier Science Ltd. All rights reserved
`
`Keywords: Lossless image compression; Medical images; Image databases; PACS
`
`1. Introduction
`
`Digital storing of medical image data is problematic,
`especially because of the requirement to preserve the best
`possible image quality. In several countries this has led to
`the interpretation that medical images should be totally
`reproducible. The reason for this is the uncertainty in the
`lossy image compression methods’ ability to preserve the
`relevant information. During the image generation phase it
`is still not known which details will later turn out to be
`important, considering the safety of the patients. The lossy
`image compression techniques, which reproduce com-
`pletely acceptable decoded images on scenes such as houses
`or landscapes, do not fulfil the strictest quality requirements
`needed in medical applications. There is always the possi-
`bility that a vague detail might give a reason to suspect some
`critical changes in a patient’s condition. For this reason, the
`lossy techniques, which tend to give high compression
`ratios, such as 1:10–1:30, are not acceptable in medical
`image compression. On the other hand, there are several
`well-known reasons why savings in the size of the image
`archives are necessary:
`• Digital medical image databases are very large; usually
`many images are taken of each patient.
`• Even if the number of patients were not radically
`increasing, the patient databases tend to grow because
`
`* Corresponding author. E-mail: olli.nevalainen@utu.fi
`
`0895-6111/98/$19.00 q 1998 Elsevier Science Ltd. All rights reserved
`PII: S0895-6111(98)00042-1
`
`•
`
`of the demand for very long storage periods of patient
`data.
`Individual patient records often consist of large images
`because of the high resolution and many bits per pixel
`used.
`• The transfer time of digital images depends on the size
`of the compressed images. The practical usefulness of a
`picture archiving and communication system (PACS)
`presupposes
`that
`transfer
`operations must
`fulfil
`reasonable time requirements.
`
`In many cases the only possibility is to trust in the lossless
`image compression methods. Unfortunately,
`their com-
`pression ratios are far from those of lossy methods, thus
`the algorithm development is severely behind the hopes
`set on the algorithms.
`This research is part of a larger research project [1,2],
`where the aim is to create a functional hospital information
`record and search system. Here we try to evaluate the exist-
`ing lossless compression methods when applied to a sample
`of the image archive. We measure the compression result by
`the compression ratio
`
`R ¼
`
`original image size
`compressed image size
`
`(1)
`
`Thus, larger values of R indicate a better compression
`method. It should be noted that the original image size
`used in the above formula contains no overhead originating
`from the header information or the byte border alignments.
`
`Netflix 1012 - Page 1
`
`

`
`324
`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`Problems in the byte border alignment occur if the number
`of bits for a pixel is not a multiple of the byte length. (For
`example, if the pixels are expressed by 12 bits, image
`formats which do not include any compression generally
`do not utilize the four last bits of the second byte.) We
`thus calculate the image size simply by the formula
`
`original image size
`¼ Xdimension p Ydimension p bits per pixel
`
`On the other hand, the compressed image size includes all
`the header information needed in the reconstruction of the
`original image data. However, this does not have a great
`effect on the compression ratios, except for some very
`small images (for which we also apply an improved format
`of storage). Note that the images are originally given in the
`DICOM format [3], which includes patient data headers. We
`have discarded these headers because they are irrelevant
`from the compression point of view.
`Fast performance is a desirable property for an algorithm
`and in some cases a long running time may even rule out the
`application of a space-efficient compression technique.
`This is why we measured the running times needed for
`compressing and decompressing the test images.
`This paper is organized as follows. In Section 2 we
`describe the image material we have used and give some
`characterizations of the image types. In Section 3 we recall
`briefly the compression techniques used. These consist of
`seven specific grayscale image compression techniques,
`four compressed image formats and 18 general all-purpose
`compressing techniques. Because the latter group contains
`numerous freely available programs, we have included only
`
`methods that are either known to compress some data very
`well or are widely used. Section 4 contains the summaries of
`the experiments and conclusions are drawn in Section 5.
`
`2. Image types
`
`the Albert Szent-Gyo¨rgyi
`The medical database of
`Medical University contains images of five different
`modalities: computed radiology (CR), computed tom-
`ography (CT), magnetic resonance (MR), nuclear medicine
`(NM) and ultrasound (US). The images originate from 10
`types of patient studies. There are altogether 125 patient
`studies, which contain in total 3147 monochrome images.
`Table 1 shows a summary of the image types and their
`properties. We observe that the pixel depths are 8, 12, 13
`or 16 bpp and a particular image type may appear with many
`different
`resolutions. The heterogeneity of
`the image
`material is added by the fact that some of the image types
`are of constant size whereas some others may appear in
`several different sizes. Appendix A contains an example
`of each image type. Three of the image types are small
`(64 3 64 or 128 3 128) and their compression presupposes
`special care for certain methods. It turns out that a simple
`compression technique (such as FELICS) works well for
`these small images whereas certain advanced techniques
`(e.g. SPIHT) do not operate for them. One solution is to
`form a larger image composed of several small images
`and handle it as a whole, see Table 1, the merged image
`types. A merged image is composed of the images of one
`study. Thus, the images to be merged are of the same size
`and type.
`
`Table 1
`Image types a
`
`Mod
`
`Type
`
`CR
`CT
`CT
`MR
`MR
`NM
`
`NM
`
`NM
`
`NM
`US
`Total
`
`PED CHEST
`ABDOMEN
`HEAD
`ABDOMEN
`SPINE
`BRAIN
`merged
`HB
`merged
`HEART
`merged
`WB
`
`merged
`
`12
`
`3
`3
`3
`3
`3
`
`bpp
`
`8
`
`3
`3
`
`3
`3
`
`Resolution
`
`16
`
`64 2
`
`128 2
`
`256 2
`
`512 2
`
`Other
`
`Study
`count
`
`Image
`count
`
`Filesize/b
`
`Real size/b
`
`3
`3
`3
`3
`
`3
`3
`3
`3
`
`3
`
`3
`
`3
`
`3
`3
`
`3
`
`3
`3
`
`3
`
`3
`3
`3
`
`3
`
`3
`
`3
`3
`3
`
`4
`17
`14
`18
`14
`4
`4
`15
`15
`17
`17
`14
`8
`125
`125
`
`4
`237
`1138
`24
`152
`480
`4
`480
`16
`510
`17
`84
`38
`3147
`1714
`
`35958856
`126134740
`548073169
`3146112
`44304888
`7871520
`7864388
`3939360
`3932432
`4185570
`4178209
`8258810
`11405914
`793278939
`793257518
`
`26969088
`100365346
`411762116
`2359296
`43155456
`7864320
`7864320
`3932160
`3932160
`4177920
`4177920
`8257536
`11405344
`620248582
`620248582
`
`aThe resolutions of the image types are marked by ‘3’. Image count gives the total number of images in all the studies. Filesize includes the space for the
`headers and the extra bits due to the byte border alignment; real size consists of the pixel data only. PED CHEST ¼ pediatric chest, HB ¼ heart and body,
`WB ¼ whole body.
`
`Netflix 1012 - Page 2
`
`

`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`325
`
`Table 2
`Lossless image compression methods a
`
`From external sources
`
`Method
`
`BTPC
`CALIC
`GIF
`JPEG-LS
`
`Program
`
`BTPC
`CALIC
`PBMPLUS
`LOCO
`
`PNG
`
`PNGLIB
`
`SPIHT
`TMW
`
`SPIHT
`TMW
`
`Ver.
`
`4
`
`0.85
`
`0.96
`
`8.01
`0.51
`
`Own implementations
`
`Method
`
`CLIC
`CPM
`FELICS
`LJPEG
`
`Options
`
`(default)
`
`(default)
`5 (predictor)
`
`Restrictions
`
`w ¼ h ¼ 2 k
`
`Author
`
`John Robinson
`Xiaolin Wu
`Jef Poskanzer
`Hewlett-Packard
`Lab
`Andreas Dilger
`
`Amir Said
`Bernd Meyer
`
`Year
`
`1997
`1995
`1991
`1997
`
`1997
`
`1995
`1997
`
`Status
`
`Restrictions
`
`Notes
`
`free
`eval
`free
`eval
`
`free
`
`eval
`eval
`
`max 8 bpp
`
`max 8 bpp
`
`w,h . ¼ 128
`
`not optimized
`
`main program for
`the library was
`written
`
`decoding of 16 bit
`images not
`successful
`
`aVer. ¼ version number, free ¼ freeware, eval ¼ for test or evaluation purposes, w ¼ image width, h ¼ image height. Default options for lossless compression
`were used except for PNG, for which the PNG_FILTER_PAETH option was used.
`
`3. Compression methods
`
`3.1. Grayscale image compression methods
`
`Table 2 contains a summary of the grayscale image com-
`pression methods used. Some of these methods are tested
`with freely available programs and others we have
`implemented according to the algorithm description from
`the literature. Our own implementations do not include
`any special code optimizations for the running time; they
`are written according to good programming style.
`CALIC, a Context Based, Adaptive, Lossless Image
`Codec [4] is a compression technique based on the pixel
`context of the present pixel to be coded (i.e. the setting of
`the pixels of some predetermined pattern of neighbour
`pixels). The method is capable of learning from the errors
`made in the previous predictions and in this way it can
`improve its prediction adaptively when the compression
`proceeds. Pixel values are predicted by a nonlinear pre-
`dictor, which selects the prediction pixels and the particular
`prediction function among several possible prediction func-
`tions on the basis of the local context. The context is built up
`of the local gradient magnitudes in horizontal and vertical
`directions. The accuracy of the prediction is improved by
`adding an estimate of the error to the predicted value. This
`estimate is the average error of the previous prediction
`values in the present context. The context for error estima-
`tion is selected in such a way that it models the magnitudes
`of the local gradient and the two previous error values both
`
`texture of the image and the
`in relation to the local
`prediction value in the most effective way.
`Four coefficients are used to weight the horizontal and
`vertical gradient magnitudes and the previous prediction
`errors, when calculating the context for error estimation.
`The coefficients should be selected on the basis of the train-
`ing set drawn from the type of images to be compressed. The
`final set of prediction errors is coded by arithmetic or
`Huffman coding.
`The coefficients used in this study have been optimized
`for ISO images and it is therefore possible that a case
`specific optimization could still improve our result. The
`final coding was performed by the arithmetic code.
`The Central Prediction Method (CPM) [5] is an iterative
`compression technique that applies the hierarchical four-bit
`coding to the image to code the prediction errors. The pixel
`values are predicted in two phases. In the first phase the
`value of a pixel is predicted as an average of the four pixels
`of the vertical and horizontal neighbours. In the second
`phase half of the remaining pixels are predicted by the
`pixel values in the two diagonal directions. A quarter of
`the pixel values remains unpredicted and will be moved to
`a separate block, which will then be compressed recursively
`by the same prediction method. Fig. 1 illustrates this.
`The recursion terminates when the size of the block of
`original pixel values is 2 3 2. The bit levels of the resulting
`image are coded by the hierarchical four-bit coding scheme.
`The coding scheme divides the bitplane of size N 3 N into
`four subplanes of size N/2 3 N/2, each of which is coded by
`
`Netflix 1012 - Page 3
`
`

`
`326
`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`two children at a lower level of the pyramid. Thus, error
`values form a binary tree. The actual coding of the image
`pyramid is done by adaptive Huffman coding. If all the
`descendants of a zero error value are also zeros, the subtree
`can be coded by a single codeword. The implementation
`used in this study works for 8-bit images only.
`Portable Network Graphics (PNG) [7,8] is essentially a
`file format for lossless compression and storage of images. It
`was developed to replace GIF. In addition to the images, one
`can also store other data in PNG files. The system accepts
`pixel depths 1, 2, 4, 8 and 16 and consumes no extra bits
`between two pixels in the storage. The compression
`method can also handle other pixel depths given to it. It is
`possible to filter the data before compression to get a better
`compression ratio.
`The actual compression is performed by the so-called
`deflate/inflate-compression. Deflate is an LZ77-based
`method which is used in several general compression
`programs (e.g. zip, gzip). Huffman coding is included in
`the method to encode the data. We used the so-called
`Paeth-filtering [9] for the image preprocessing. The com-
`pression level option was set to default compression (6),
`which turned out to be two to three times faster than the
`best compression with only 1% decrease of compression
`ratio. A simple main program was made to call the png
`library, which is freely available on the Internet.
`The SPIHT-compression technique (Set Partitioning in
`Hierarchical Trees) [10] transforms the image to a multi-
`resolution representation by the wavelet transform or, in the
`case of lossless compression, by the S transform. This trans-
`formation is similar to the subband decomposition, but uses
`only integer operations. The S transform scans the columns
`of the image and calculates for each successive pair of the
`pixels the average and the difference. The averages are
`stored at the upper part of the image (l), and the differences
`are stored at the lower part (h), see Fig. 3. The same is
`repeated for the columns of l and h parts giving ll, lh, hl
`and hh images. At the next level
`the ll block will be
`processed in the same way. This gives a multiresolution
`pyramid of the image with reduced variance.
`
`Fig. 1. Prediction scheme of CPM.
`
`a bit in a four-bit codeword. If all the pixel values in a
`subplane are zero, the subplane is coded by a zero. Other-
`wise it is coded by a one. The procedure is recursively
`applied to each of the subplanes. The recursion stops
`when the subplane is coded by a zero or when the size of
`the subplane is 1 3 1. The four-bit method presupposes a
`square-like form with 2 k pixels in each side.
`is
`[6]
`The Binary Tree Predictive Coding (BTPC)
`designed for lossy and lossless image compression and is
`suitable for both photographs and graphics. The method
`forms a binary pyramid of the image by recursively dividing
`the image into two subimages whose sizes are half of the
`size of the previous image. The first subimage (S) is formed
`from the original image by leaving out alternate pixels; from
`each odd-numbered row the even-numbered pixels are
`discarded, and from each even-numbered row the odd-
`numbered pixels are discarded, see Fig. 2. The second
`subimage (E) stores the pixels missing from the first
`subimage, but in a transformed form; a context model
`(shape-based prediction, which uses different predictors
`for edges, flat areas, etc.) uses the neighbouring pixels in
`the subimage S to predict a pixel value. The actual stored
`value is the prediction error, i.e. the difference between the
`predicted and the real pixel value. The prediction error
`transform aims to a skewed distribution of values, which
`can be effectively compressed. The next
`layer of the
`image pyramid is constructed from the previous S subimage
`by the same process.
`Each error value of a particular level of the pyramid has
`
`Fig. 2. Principle of the pyramid coding in BTPC.
`
`Netflix 1012 - Page 4
`
`

`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`327
`
`Fig. 3. Principle of the SPIHT compression technique.
`
`After each one-dimensional transform, the pixel differ-
`ences are replaced by the difference of transformed pixel
`value and prediction value. This is known as the P trans-
`form. The prediction values are calculated as a function of
`pixels of the lower levels of the pyramid and the known
`pixels of the current level. The prediction function can be
`selected on the basis of the properties of the image, but the
`coefficients of the function are not very critical. Arithmetic
`coding concludes the encoding process.
`TMW [11] is based on the use of linear predictors and
`implicit segmentation. The compression process is split into
`an analysis step and a coding step. In the first step, linear
`predictors and other parameters suitable for the image are
`calculated. These are included in the compressed file and
`subsequently used for the coding step.
`In TMW, three different kinds of predictors are used.
`Pixel-predictors predict a pixel value. Sigma-predictors
`predict the magnitude of pixel-predictor’s prediction error.
`Blending-predictors predict how well suited a particular
`pixel-predictor is to predict a pixel value. These all are
`based on the corresponding values of the causal neighbours.
`In the analysis step, the parameters of these predictors are
`iteratively optimized. This is very time-consuming. How-
`ever, if we have several images that resemble one another, it
`is sufficient to build a model for only one of them. This
`model can then be used for compressing all the other similar
`images. The coding step is much faster and basically con-
`sists of arithmetic coding of the prediction error according
`to the distribution.
`The implementation used by us failed to reconstruct the
`16-bit images. However, it seems that the problem lies in the
`decompressor. Thus, the results have been presented. We
`used the parameter file created for the test image ref12b-0 in
`all our tests, because the model generation software was not
`available. Better results would probably have been achieved
`if we had been able to build a model for each image type.
`The Low Complexity Lossless Compression for Images
`(LOCO-I) [12] is based on context modelling and predictive
`coding combined with Huffman coding. The pixel values
`are predicted adaptively according to the output of a simple
`edge detector. The prediction value is a simple function of
`the values of the predictor pixels, which are chosen among
`three neighbouring pixels.
`The prediction residuals are coded by the Golomb–Rice
`code [13,14] in the context that is built out of the differences
`of four local pixel values. The differences are quantized to
`reduce the number of contexts. Golomb–Rice codes are
`characterized by one parameter that forms the optimal
`
`probability distribution for the code. For the present context,
`the parameter value, which is intended to yield as short a
`code length as possible, is calculated from the estimate of
`the expected prediction error magnitude in the context. For
`flat regions, which are detected by checking the context,
`LOCO-I uses run-length coding. Here, a run of symbol is
`coded by coding the length of the run.
`LOCO-I has been chosen as the algorithm of JPEG-LS,
`the emerging lossless/near-lossless compression standard
`for continuous-tone images. A freely available implementa-
`tion of the method was used for testing. Default values were
`chosen for the thresholds for context quantization.
`The JPEG (Joint Photographics Experts Group) standard
`for lossless compression LJPEG [15] is based on a predic-
`tion method combined with the entropy coding. The
`prediction method transforms the original
`image to an
`error
`image with reduced variance using a predictor,
`which is initially chosen among eight possible predictors.
`The prediction is a function of some of the three previous
`neighbouring pixels (pixels to the left, above and upper-left
`of the current pixels). The error image is finally compressed
`using Huffman or arithmetic coding.
`In this study the Huffman code was applied for the com-
`pression of the error images. The Huffman codes were
`derived from static codes for 8-bit images by adding the
`missing codewords. The Huffman code chooses a category
`for the error value on the basis of the most significant bit.
`The codeword for the error value is the codeword for the
`category combined with rest of the bits in the error value
`(the sign and the least significant bits). In the tests we have
`used predictor W þ b(N ¹ NW)/2c (W corresponds to the
`pixel west of the current pixel and so on).
`FELICS (Fast, Efficient, Lossless Image Compression
`System) [16] is based on context modelling and Rice coding.
`The context for a pixel consists of the two previous pixel
`values L and H, which make up a value range from the
`smaller (L) of the values up to the larger (H). Coding of
`the present pixel value P starts by checking whether the
`pixel value belongs to the value range or not. The result
`of this check is indicated by the first bit in the codeword.
`If the value belongs to the range, the difference P ¹ L is
`coded using a binary code. If the pixel value does not belong
`to the range, the second bit in the codeword indicates
`whether the value is bigger or smaller than any of the values
`in the range. In the first case the difference P ¹ H ¹ 1 is
`coded using the exponential Rice prefix code. Otherwise,
`the difference L ¹ P ¹ 1 is coded using the same code. The
`Rice code has one parameter. The parameter value giving
`
`Netflix 1012 - Page 5
`
`

`
`328
`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`the shortest code length is estimated for each context during
`the coding process.
`Two implementations of FELICS were used in this study.
`One program was implemented for images with pixel depth
`8 while the other was implemented for images with pixel
`depths from 9 to 16.
`The Context-based Lossless Image Compression (CLIC)
`[17] uses several contexts to estimate probability density
`functions of error values, which are finally coded by the
`arithmetic coding. The error values are calculated as a
`difference between a pixel value and a prediction. The
`prediction function can be chosen among several possible
`functions such as (N þ W)/2, where N corresponds to the
`pixel value north (above) of the current pixel, etc. This is the
`prediction function we used. The probability density func-
`tion of the error values is estimated for each context by
`
`calculating frequencies of error values. In order to save
`memory, this can be done bitwise by calculating the fre-
`quency of each bit (having value one) in each context. In
`this case each bit is coded separately. The context was built
`out of the error values of the pixels above and left of the
`current pixel.
`We used a backup model when a context of the main
`model was updated less than 64 times. The prediction func-
`tion of the backup model was the same as in the main model.
`The context of the backup model was the error value of the
`pixel left of the current pixel.
`The Graphics Interchange Format (GIF) is a widely used
`image format by CompuServe Incorporated. It applies an
`LZW based compression method. The format is restricted to
`an 8-bit grayscale and palette colour images. The format is the
`most suitable for storing drawn images with few colours.
`
`Table 3
`General-purpose compression programs a
`
`Name
`
`ACB
`ARJ
`BZIP
`BZIP2
`ESP
`GZIP
`HA
`JAR
`PKZIP
`Quantum
`RAR
`RKIVE
`SZIP
`SZIP
`UC2
`UFA
`X1
`YAC
`
`Name
`
`ACB
`ARJ
`BZIP
`BZIP2
`ESP
`GZIP
`HA
`JAR
`PKZIP
`Quantum
`RAR
`RKIVE
`SZIP 1.04
`SZIP 1.05Xd
`UC2
`UFA
`X1
`YAC
`
`Version
`
`2.00c
`2.60
`0.21
`0.1pl2
`1.92
`1.2.4
`9.999b
`1.02
`2.04g
`0.97
`2.02
`1.90b1
`1.04
`1.05Xd
`3.0 Pro
`0.034b4
`9.95a
`1.02
`
`Author
`
`George Buyanovsky
`Robert Jung
`Julian Seward
`Julian Seward
`GyikSoft
`Jean-loup Gailly
`Harri Hirvola
`Robert Jung
`PKWARE Inc.
`Cinematronics
`Eugene Roshal
`Malcolm Taylor
`Michael Schindler
`Michael Schindler
`AIPNL
`Igor Pavlov
`Stig Valentini
`Aleksandras Surna
`
`Options used
`
`Year
`
`1997
`1995
`1996
`1997
`1997
`1993
`1995
`1997
`1993
`1995
`1997
`1997
`1997
`1997
`1995
`1997
`1997
`1995
`
`Status
`
`DOS
`
`W95
`
`Linux
`
`src
`
`beta
`
`solid
`
`Shareware
`Shareware
`Freeware
`Freeware
`Freeware
`Freeware
`Freeware
`Shareware
`Shareware
`Freeware
`Shareware
`Shareware
`Freeware
`Freeware
`Shareware
`Shareware
`Freeware
`Freeware
`
`3
`3
`
`3
`3
`3
`3
`3
`3
`3
`3
`3
`3
`3
`
`3
`3
`
`3
`3
`3
`
`3
`3
`
`3
`3
`3
`3
`3
`3
`3
`
`3
`3
`
`3
`3
`
`3
`
`3
`3
`
`3
`
`3
`3
`
`3
`3
`
`3
`
`3
`
`3
`
`3
`3
`
`3
`
`3
`3
`3
`3
`
`3
`
`3
`3
`3
`3
`3
`3
`3
`3
`
`u (max)
`-jm (max)
`-9 (max block size)
`-9 (max block size)
`/mm (8-bit multimedia)
`-9 (max)
`21 (choose best)
`-m4 (max)
`-ex (max)
`-c5 (max) -t21 (max table size)
`-mm (multimedia) -m5 (max) -s (solid)
`-mtx (best mode) -mm2 (multimedia forced)
`
`-b41 (max blocksize) -o255 (max order) -r1 (recordsize)
`-tst (tight multimedia)
`-m5 (text compression) -mx (max) -s (solid)
`u (solid) m2 (method 2)
`
`Notes
`
`16 MB RAM required
`many options, popular
`single file compressor
`no arithmetic compression
`
`single file compressor, popular
`
`popular
`
`popular
`-mtx requires 32MB RAM
`single file compressor
`new, experimental version
`
`aDOS ¼ program is available for plain DOS, W95 ¼ program (possibly a separate version) supports Windows 95y long file names, Linux ¼ Linux version is
`available, src ¼ program sources are available, beta ¼ program is in beta testing, i.e. not an official release, solid ¼ program supports solid compression.
`
`Netflix 1012 - Page 6
`
`

`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`329
`
`3.2. General-purpose lossless compression methods
`
`Table 3 contains a summary of the general-purpose
`lossless compression methods applied in this work. Most
`of these programs are selected according to our previous
`knowledge on their ability to compress general data.
`GZIP, PKZIP and ARJ have been included because they
`are widely used.
`It should be noted that BZIPs, GZIP and SZIPs are single
`file compressors. When several images were compressed in
`the same file, TAR program was used in these cases. Thus,
`compression was solid, i.e. all the files to be compressed are
`handled as a single file. It is well known that if the files are
`of the same kind, this usually gives better results than
`compressing the files separately. We have tested solid com-
`pression if the program supports it. Note that TAR inserts
`headers between the files, but this was found to have no
`practical effect on the compression ratio.
`There are two versions of BZIP and SZIP. BZIP2 0.1pl2
`is a newer version of BZIP 0.21, but its compression results
`are worse. This is because instead of the arithmetic coding
`used in BZIP, Huffman coding is used in BZIP2. (On the
`other hand, BZIP2 is patent-free, unlike BZIP.) SZIP
`1.05Xd is an experimental, non-supported version of
`SZIP. It has some options so that it can be made to compress
`better at the expense of increased running time.
`
`4. Empirical tests
`
`Table 4 shows a summary of the test results for the 11
`image compression techniques. The numbers shown in the
`table are the average compression ratios calculated for the
`different pixel depths and sizes. Here we will express the
`results as averages over all these. For small images we show
`(if possible) the results for both single and merged images.
`The results of Table 4 indicate that the four most space-
`efficient compression techniques for this set of data are
`
`CALIC, TMW, SPIHT and JPEG-LS. The table also
`shows the average compression ratio of each compression
`method. Here the average is calculated as the total size of
`uncompressed images divided by the size of compressed
`database. Therefore, the averages are close to the results
`for CT/HEAD type of images, as two-thirds of the database
`is of this type. (Note that the weighted average is not the
`only way to compare the different compression tech-
`niques—compare with the unweighted average of all
`compression ratios of images or studies.) The average com-
`pression ratio of CALIC is 2.98 for all image types. For
`SPIHT and JPEG-LS the results are essentially of the
`same quality. The average compression ratio of TMW was
`as much as 3.17. This is because TMW compressed CT/
`HEAD images over 10% better than CALIC. CALIC
`compressed most of the other images better. However, if a
`unique model were generated for each image type, the
`compression results of TMW would probably be better.
`The variation of the compression ratios of the techniques
`is large among different image types. For CALIC the com-
`pression ratio ranges from 2.06 to 5.23. For single images
`the range is even greater, 1.52–24.16 (there are three images
`with compression ratios . 10). In addition, there are some
`cases where the relative order of the three most effective
`techniques is not clear; for MR images SPIHT gives the
`highest compression ratios and for NM/BRAIN images
`JPEG-LS has the highest ratio. It is interesting to note that
`FELICS is essentially as efficient as CALIC for the very
`small (64 3 64) images. Also LJPEG works relatively well
`for
`the very small
`images.
`JPEG-LS is especially
`space-efficient for the 8 bpp images.
`In the next category of the techniques we have FELICS
`and CLIC. FELICS succeeds well on the very small
`images—it gives the best results for the NM/HB images
`compressed separately. This is probably because FELICS
`is a rather simple method and the very small
`images
`(64 3 64) are not suitable for more sophisticated methods.
`For example, the results of TMW are very poor on the small
`
`Table 4
`The average compression ratios for the image compression methods by image type (JP-LS ¼ JPEG-LS, FEL ¼ FELICS)
`
`Mod
`
`Type
`
`TMW
`
`CALIC
`
`SPIHT
`
`JP-LS
`
`CLIC
`
`CR
`CT
`CT
`MR
`MR
`NM
`
`NM
`
`NM
`
`NM
`US
`Total
`
`PED CHEST
`ABDOMEN
`HEAD
`ABDOMEN
`SPINE
`BRAIN
`merged
`HB
`merged
`HEART
`merged
`WB
`(missing)
`
`merged
`
`2.71
`3.77
`3.16
`1.98
`2.81
`3.35
`5.04
`2.08
`3.45
`2.11
`3.54
`3.81
`2.80
`3.17
`3.21
`
`2.86
`3.72
`2.83
`2.06
`2.90
`4.87
`5.23
`3.28
`3.44
`3.36
`3.54
`4.09
`3.02
`2.98
`2.99
`
`2.81
`3.44
`2.82
`2.03
`2.94
`4.58
`4.78
`
`3.39
`
`3.48
`3.84
`2.54
`
`2.94
`
`2.74
`3.59
`2.65
`2.02
`2.81
`4.96
`5.11
`2.71
`3.39
`2.77
`3.47
`4.05
`2.90
`2.81
`2.82
`
`2.68
`3.40
`2.50
`1.86
`2.69
`4.52
`4.81
`2.44
`3.26
`2.51
`3.35
`3.82
`2.73
`2.67
`2.68
`
`FEL
`
`2.60
`2.90
`2.18
`1.89
`2.67
`3.67
`3.77
`3.28
`3.36
`3.36
`3.45
`3.28
`2.40
`2.36
`2.36
`
`LJPEG
`
`PNG
`
`CPM
`
`BTPC
`
`GIF
`
`2.24
`2.68
`2.05
`1.71
`2.42
`2.55
`2.55
`2.90
`2.91
`2.96
`2.97
`2.33
`2.02
`2.18
`2.18
`
`1.97
`2.44
`1.74
`1.45
`2.05
`3.79
`3.99
`2.31
`2.46
`2.36
`2.51
`3.24
`2.50
`1.90
`1.90
`
`1.57
`2.19
`3.28
`
`2.58
`
`2.64
`
`4.30
`4.83
`
`4.04
`4.40
`
`3.77
`2.40
`
`3.19
`1.94
`
`Netflix 1012 - Page 7
`
`

`
`330
`
`J. Kivija¨rvi et al./Computerized Medical Imaging and Graphics 22 (1998) 323–339
`
`Fig. 4. Average compression times (s) for image compression methods with respect to filesize (bytes) for Axil320. The compression times of TMW, BTPC,
`GIF, FELICS and LJPEG are omitted. TMW results were run on a different hardware. BTPC and GIF were not able to compress 16 bit images. FELICS
`compression times are essentially the same as for SPIHT. LJPEG compression times are between JPEG-LS and CALIC.
`
`images because TMW generates a 1596 byte parameter file
`for each image, which we have included in the compressed
`filesize.
`The CPM algorithm presupposes a square-like image and
`we therefore restricted our test with it to this kind of images
`only. Similarly, our GIF results are restricted to 8 bpp
`images. For both of these techniques the results do not
`encourage modifications allowing other image types. Also
`the implementation of BTPC was restricted to 8 bpp images.
`However, moderately good results were obtained by BTPC.
`One aspect is the merging of smaller images to form a
`larger one. A general observation is that the mutual prefer-
`ence of the different methods is preserved by the merging
`operation. The compression ratio, however,
`increases
`radically (for certain images even by 30%). This is because
`the information obtained from one image can be utilized
`when compressing other
`images. The improvement
`is
`greatest for JPEG-LS and CLIC. See Appendix B for
`more details on the test results.
`Tables 6–8 give a summary of the compression and
`decompress