Proceedings of the 2005 IEEE
Engineering in Medicine and Biology 27th Annual Conference
Shanghai, China, September 1-4, 2005
An Efficient Medical Image Compression Scheme
Xiaofeng Li, Yi Shenഘ Jiachen Ma
Department of Control Science and Engineering, Harbin Institute of Technology, Harbin, 150001, China
Abstract—In this paper, a fast lossless compression
scheme is presented for the medical image . This
scheme consists of two stages. In the first stage, a
Differential Pulse Code Modulation (DPCM) is used to
decorrelate the raw image data, therefore increasing
the compressibility of the medical image. In the second
stage, an effective scheme based on the Huffman coding
method is developed to encode the residual image. This
newly proposed scheme could reduce the cost for the
Huffman coding table while achieving high compression
ratio. With this algorithm, a compression ratio higher
than that of the lossless JPEG method for image can
be obtained. At the same time, this method is quicker
than the lossless JPEG2000. In other words, the newly
proposed algorithm provides a good means for lossless
medical image compression.
Keywords—Lossless image compression, Huffman
coding, medical image
I. INTRODUCTION
The trend in medical imaging is increasingly toward
direct digital image data acquisition. Currently, most
modalities such as computed tomography (CT), magnetic
resonance Imaging (MRI), positron emission tomography
(PET), single photon emission computed tomography
(SPECT), produce images directly in digital form. Digital
X-ray imaging systems are also being introduced for chest
and breast imaging, and various types of film digitizers
are being used to convert traditional X-ray films into digital
format. The main advantages of digital imaging lie in its
ready support of image transfer and archival, and the
possibility of manipulating and enhancing diagnostic
information. However, digital image transmission and
storage are facing major challenges due to the growing size
of medical image datasets as well as practical limitations in
transmission bandwidth and storage space. This has led to a
search for effective image compression methods[1].
During the past two decades, various lossless and lossy
image coding techniques have been developed. Although
lossy coding methods can achieve high compression ratios,
e.g., 50:1 or more, they lack the ability to exactly recover
the original data. While a lossless coding method will
allow exact recovery from the compressed version of the
data, the compression ratio achievable is usually only of
the order of 2:1 or 3:1. In medical applications, lossless
coding methods are required since loss of any information
is usually unacceptable.
Most of the recent advances in image coding have
taken place in the area of lossy compression, but lossless
0-7803-8740-6/05/$20.00 ©2005 IEEE. 3437
coding algorithms have not seen much progress. Therefore,
though lossless JPEG having been used for many years, it’s
still been widely used in medical image compression and
been specified as medical image compression criterion in
DICOM (Digital Imaging and Communications in
Medicine)[2] because of it’s simpleness and efficiency.
Lossless JPEG have two versions: DPCP with arithmetic
coding and DPCM with Huffman coding. The arithmetic
version typically provides small compression
improvements over the Huffman coding, but at a much
higher computational cost[3].
In this paper, we present a simple algorithm
incorporating DPCM and a Huffman encoding of most-
likely residual magnitudes. We then compare it with the
DPCM-Huffman method of lossless JPEG and lossless
JPEG2000.
The rest of this paper is organized as follows. In
section II, we shall review the Huffman coding scheme. In
section III, the improved method will be proposed. In
section IV some experimental results shall show that our
newly proposed algorithm indeed produces good image
quality. In section V, the discussion and conclusion shall be
given in section.
II. THE DPCM AND HUFFMAN CODING IN JPEG
Differential Pulse Code Modulation (DPCM) is
transformation for increasing the compressibility of an
image. It consists of scanning the image and predicting the
next pixel’s value. There are several modes to predict the
next pixel’s value. In this paper we use the average of the
pixels to the left and the pixel above as the predicted value.
In a lossless coding scheme, the original image is
reproducible by storing the pixels at the beginning of the
scan and decoding in the same scan order.
The set of differences between each pixel and its
predicted value is the residual image. The residual
distribution is typically zero-mean and much more compact
than the distribution of the original image. A more compact
distribution results in lower entropy which determines the
minimum average codeword length that is attainable
without information loss.
Fig.1. shows vessel image of 512×512×8bpp, Fig.2.
show the DPCM residual.
The Huffman coding technology is a commonly used
scheme for data compression because it is very simple and
effective. It requires the statistical information about the

Fig.1. Vessel image.
Fig.2. The DPCM residual of Fig.1.
distribution of the data to been coded. Besides an identical
coding table is used in both the encoder and the decoder.
The details of the Huffman coding technique is described
in the following.
1) Histogram the residual image.
2) Build a coding tree by sorting the histogram and
successively combining the two bins of lowest value until
one bin remains.
3) Encode the residual image and save the coding tree
with the encoded values.
In Step 1, the Huffman tree of the given data is built by
successively combining the two bins with the lowest value.
The Huffman coding table is then constructed to record the
mapping of the data of each bin and the corresponding
encoding bit pattern. The performance of the Huffman
coding technique relies on the distribution of the input data.
If the input data has a large dynamic range, the Huffman
coding technique requires a higher computation cost.
III. THE IMPROVED SCHEME
The Huffman coding of the most-likely magnitude is
proposed to solve the problem expounded in section IIG
since the fundamental goal of most-likely magnitude[4] is
to encode the residual image generated for the linear
predictor. It exploits two properties of image residual
distribution:
1) The distribution of these residual magnitudes is
symmetrical to the origin.
2) The number of the small residual magnitudes is
much larger than that of the large residual magnitudes.
3438
By exploiting Property 1, the most-likely magnitude
scheme cuts down the number of code calculations by half
by creating Huffman codes only for the residual
magnitudes and then adds a sign bit. By exploiting
Property 2, the most-likely magnitude scheme uses the
Huffman coding technique only for the most-likely residual
magnitudes and employs another special prefix code to
identify the least-likely residual values. The details of the
most-likely magnitude scheme are described in the
following.
Step 1. Compute the residual histogram H, such that
H(x) =the number of pixels having residual magnitudes.
Fig.3 shows the residual magnitude histogram of Fig.2.
Step 2. Compute the symmetry histogram S such that
S(y) = H(y) + H(-y) y>0 (1)
Step 3. Find the range threshold R such that
∑ S ( j ) ≤P × N < ∑ j = 0 S ( j )
R −1 R
j =0
(2)
Where N is the total number of the pixels in the input
residual image and P is the desired proportion of the most-
likely magnitudes. Note that 0 ≤ P ≤ 1 and 0 ≤ R .
Step 4. Construct the modified symmetry histogram S1
with R+1 entries, where
⎧⎪ S ( y ) 0 ≤ y ≤ R-1
S1 ( y ) = ⎨ (3)
⎪⎩ N − ∑ j = 0 S ( j ) y = R
R −1
Note that S1(R) stores the total number of the least-
likely occurrences.
Step 5. Construct the Huffman tree and generate the
Huffman table C for histogram S 1 using the traditional
Huffman coding technique.
Step 6. Encode each residual magnitude x by the
following three rules.
Rule 1: if 0”x”R-1 then output conc(C(x), 0).
Rule 2: if -(R-1)) ”x<0 then output conc(C(|x|), 1).
Rule 3: if x•R or x”-R then output conc(C(R), x).
Here the function conc(x, z) concatenates the bit
representations of x and z. Here C(x) denotes the encoded
bit pattern for input x stored in the Huffman coding table.
x 10
4
2.5
1.5
1
0.5
0
0 50 100 150 200 250
Fig.3. The histogram of Fig.1
In Step 1), the residual histogram H is computed and to
be transformed into the symmetry histogram S in Step 2.
The range of symmetry histogram S is about half of the V. CONCLUSIONS
original residual histogram H. However, the information The most-likely magnitude approach results in an
about the sign value should be kept for residual re- algorithm with significant time and compression
production in the decoder. That is, to record the sign value improvements over the JPEG (Huffman). These advantages
of each residual, one sign bit should be added in the become most apparent for images with high residual
encoding step. Given a proportion threshold P, where dynamic range. This method is simple to implement and
0”P”1, the range threshold R to classify the most-likely enables systems to compress imagery in read-time with a
magnitude and the least-likely magnitude is decided in Step software-only implementation. From Table I and Table II
3. If P is set to 0.9, it means that we want to encode 90% of we can see that, though compression ratio is less than
the residual magnitudes using the Huffman coding JPEG2000, the most-likely magnitude approach much
technique with slight modification. Thus, a switch value more quick than lossless JPEG2000 when compression the
should be added in front of these least-likely magnitudes to same medical image.
accommodate the encoding technique. So, these least-likely
magnitudes are collected together in Step 4. At present,
there are R+1 entries in the modified symmetry histogram TABLE I
S1. COMPRESSION RATIO
After completing the computation of histogram S1, the
standard Huffman coding technique is applied to the Image Our scheme JPEG(Huffman) JPEG2000
vessel 2.20 1.83 2.48
construction of the Huffman tree and the Huffman coding pelvis 3.12 2.82 3.50
table for S1 in Step 5. The actual encoding process for input head 1.64 1.45 1.94
residual x is done in Step 6. If x is a most-likely magnitude,
the coding output is composed of the encoded bit pattern TABLE II
for C(|x|) followed by a sign bit. Otherwise, if x is a least- COMPUTATION TIME IN MSEC
likely magnitude, the coding output is composed of the
encoded bit pattern for C(R) followed by the binary Image Our scheme JPEG(Huffman) JPEG2000
vessel 52 60 411
representation of x. pelvis 598 680 3200
The parameter P provides a means of trading off head 15 20 151
speed for compression performance. The authors claim that
P=0.99 is a good compromise for a wide variety of images. REFERENCES
A large value of P generally leads to more computation
cost though a Huffman table with more entries can be [1] Liang Shen , Rangaraj ,and M. Rangayyan, “A Segmentation-
Based Lossless Image Coding Method for High-Resolution Medical
stored. On the other hand, if the parameter P is small, the Image Compression,” IEEE Trans. Medical Image, vol. 16, no. 3, pp.
most-likely magnitude scheme has some loss in encoding 301-307,June. 1997.
these least-likely magnitudes. [2] National Electrical Manufacturers Association (NEMA), Digital
Imaging and Communications in Medicine (DICOM), 2003.
[3] Gopal Lakhani, "Modified JPEG Huffman Coding," IEEE Trans.
IV. EXPERIMENT RESULT Image Processing, vol. 12, no. 2, pp.159-169, feb. 2003
[4] Yu-chen Hu and Chin-Chen Chang,, “A New Lossless Compression
All the experiments are done upon DELL LATITUDE Scheme Based On Huffman Coding Scheme For Image
D800 notebook PC with the windows 2003 OS, and use the Compression,” ELSEVIER. Signal Processing Image
Communication., vol. 16, pp. 367–372, 2000.
matlab7.0.0 and jasper-1.701.0 as the emluator tool.
The compression algorithms we consider here include:
Lossless JPEG (Huffman).
DPCM with most-likely magnitude Huffman coding
for P=0.99.
Lossless JPEG2000.
Images used in this paper:
1. 512×512 vessel images with 256 gray levels.
2. 1680×2048 X-ray image of pelvis with 256 gray
levels
3. 256 ×256 MRI image of head with 256 gray levels.
The criterion used to compare them is compression
ratio(with respect to the number of bits), Which is defined
as follows:
original file size
Compression Ration(CR)=
compressed file size

3439