Compression in Medical Images (Paper Survey)

Compression in Medical Images
(Paper Survey)
Mohammed Jirari
Fall 2002
CS 74401
Need for Image Compression in
Medical Images
• Image compression plays an important role
in telematics applications and especially in
telemedicine.
• Diagnosis is effective only when
compression techniques preserve all the
relevant info needed. (Lossless)
Lossy Compression
• More efficient in storage and transmission.
• No guarantee that characteristics needed for
medical image diagnosis are preserved.
• Image characteristics are preserved in the
coefficients of the domain space in which the
original image is transformed. (In DWT, the
wavelet coefficients keep all the info).
• Goal is to discard only the irrelevant wavelet
coefficients according to a criterion (magnitude of
values).
2 Proposed Methods
• Applying different thresholds to uniquely
defined regions of the transform domain
(opposed to the DWT where the same
threshold is applied to the whole image).
• Separately, applying the same
transformation to the regions of interest in
which the image could be divided according
to a predetermined characteristic (texture).
First Method
• The original image is transformed via the 2D
DWT into bands of wavelet coefficients.
• Fuzzy c-means clustering is applied to each band,
dividing it into 2 classes.
• Important regions have lower compression ratio
than non-important ones.
• To reconstruct, the inverse 2D DWT is applied to
the remaining wavelet coefficients (ones removed
during compression are equal to 0).
Second Method
• Cluster the image into 2 classes (significant and
non-significant textural regions).
• Texture identification analysis is performed based
on 4 cooccurrence matrices:
* Energy Angular Second Moment
f1   p (i, j ) 2
i
* Correlation
j
Ng Ng
f2 
 (i * j ) p(i, j )  
i 1 j 1
 x y
xx
y
Second Method (cont.)
* Inverse Difference Moment
1
f 3  
p(i, j )
i
j 1  (i  j )
* Entropy
f 4   p(i, j ) log( p(i, j ))
i
j
Second Method (cont.)
• When pattern is texturally significant and
cooccurrence matrices are used, then upper left
point of the corresponding sliding window takes
on label 255, otherwise 0. (for each gray level
image a new black-white image IMP results)
• Decompose the original image into 2 images G1
and G2 by:
G1(x0,y0)=MIN(G0(x0,y0),IMPG0(x0,y0))
G2(x0,y0)=MIN(G0(x0,y0),255-IMPG0(x0,y0))
Second Method (cont.)
• 2D DWT is applied to G1 and G2 (compression
ratio for DWT-G1 is 60% and DWT-G2 is 80%,
we get DWT-G1’ and DWT-G2’ respectively).
• Reconstruct the image by:
G^0=a*INV-2D-DWT-(DWT-G1’)+
b*INV-2D-DWT-(DWT-G2’)
NOTE: Main problem is the elimination of blocking
effects in the partitions boundaries (need to
smooth the reconstructed image in these
boundaries).
Region Of Interest & Colon CT
Image Compression
Segmentation of ROI
Segmenting the colon from the CT data sets
consists of three steps:
* The air is separated away from the tissue
by intensity thresholding
* The colon wall that surrounds the air is
extracted by 3D extension of Sobel’s
derivative operation.
* A morphological 3D grassfire operation
determines the colon-wall region within 5pixel margin.
ROI Based Compression Scheme
• Once the ROI is segmented in each slice
• A hybrid compression scheme is used for coding
the images.
• The first slice of the volume is compressed with a
lossless coder.
• Each slice is then coded by motion compensated
coding.
• The difference between the real image ROI block
and the predicted ROI block is coded by an
entropy minimizing lossless coder(Huffman).