MEDICAL IMAGE COMPRESSION USING REGION-OF-INTEREST VECTOR QUANTIZATION András Czihó*#$, Guy Cazuguel*#, Basel Solaiman*#, Christian Roux*# * ENST-Bretagne, Dépt. ITI, B.P. 832 , 29285 Brest Cedex - France # Laboratoire de Traitement de l'Information Médicale (LATIM), Brest - France $Technical University of Budapest, Dept IIT, Budapest XI, Muegyetem rkp. 9, 1521 - Hungary E-mail:
[email protected] Abstract In this paper an image compression approach is proposed for medical applications. The algorithm is based on the Vector Quantization and adopts the idea of Region-OfInterest. The image to be compressed is first segmented into regions and a separate codebook is used for compressing every region. The size and the number of codewords may be different in the codebooks according to the diagnostic importance of the corresponding image region. This permits to create appropriate codebooks with representative codewords, and to obtain good reconstruction quality in relevant zones, while reinforcing the compression in less important regions. The proposed approach is tested on ultrasound esophagus images and is shown to be very promising. I. Introduction Amongst lossy signal compression approaches the Vector Quantization [1] is the optimal method in the sense that by increasing the vector length and the codebook size, better performance can be obtained than using any other block coding technique. Although the rapidly growing memory and computation requirements do not permit approximate arbitrarily closely the optimal performance, VQ has been proved to be a very straightforward image compression approach[2]. For instance, the use of variable size codewords was proposed according to the quadtree decomposition of images in order to proceed with large blocks whenever it is possible [3,4,5]. VQ has the particular advantage of being able to exploit prior knowledge on the images to be compressed. Since a codebook has to be generated before compression, one has to have a training set, i.e. several images that are representative of the images to be compressed. Thus, VQ is not a "universal" approach (as e.g. JPEG may be called) since it cannot work well for images of types that differs from the ones the training set was issued from. However, for images of the same type it can work very well (better than general methods), since the codebook contains representative codewords. The use of digital images in medical field verifies the above mentioned condition. That is, every images of the same
medical domain represents the same thing: the same part of the human body, the same organ, etc. Thus, a very suitable codebook can be created and thus VQ can work very efficiently. In [4,5] we have shown that a quadtree VQ scheme works better than the JPEG compression standard in medical environment, especially at high compression rates. In the present paper we propose another VQ scheme that also uses variable size codewords. However, in this case the appropriate codebook to be applied is chosen according to prior knowledge on the diagnostic importance of every region. Thus, the proposed algorithm is a region-of-interest (ROI) approach, and aims at exploiting as much information known a priori as possible. In the next section the standard VQ compression algorithm is revisited, while Section III. presents the general scheme of the proposed method referred to as ROI-VQ. The concrete application we study is the echoendoscopy imaging of the esophagus presented in Section IV. Simulation results are reported in Section V., and the paper is closed with the conclusions. II. Vector Quantization In VQ, the image is divided into small nonoverlapping blocks. E.g. 4x4 pixel blocks are considered as vectors of dimension 16. At the encoder, each vector xi of the image is compared to the elements of a codebook W={w0,w1,...,wN-1}, called the codevectors or codewords, and only the index of the nearest codevector is transmitted. The best matching codevector is selected according to some distortion measure, which is in general the mean square error. The decoder reconstructs the signal by simply performing table-lookup operation to fetch codevectors from a codebook which is identical to that of the encoder. The encoding and decoding scheme is shown in Fig.1. ...
Original image
...
index block
Best matching search
Table look-up Reconstructed image
Codebook
Codebook
Fig. 1: VQ encoder and decoder
The codebook must contain vectors that represent well the images to be compressed. Several methods are used in constructing codebooks. They apply, in general, a learning method on the training set issued from available images which are supposed to be representative of the images to be compressed.
IV. Esophagus image compression by ROI-VQ
III. ROI Vector Quantization In medicine images are being used to a well-defined task: the analysis of different parts of the human body. Moreover, on a certain image different regions may have very different diagnostic importance, since eventual pathologies may happen only on several regions of the image corresponding to well defined organs or parts of organ. For example, on an X-ray hand image the expert is mainly interested in the states of the bones. These bones occupy finally a very restricted part of the image, therefore the rest, a large image zone may be compressed with a low bitrate since a poorer reconstruction quality is tolerable. This idea and the VQ advantages lead to the proposed Region-Of-Interest VQ (ROI-VQ) approach. In this method, a separate codebook is generated for every region (or image 'object', i.e. organ). The properties of these codebooks, i.e. the codebook size and the block size, are chosen according to the medical importance of the given object. If an object is diagnostically important, a large codebook containing small codewords is created. Inversely, for less important regions, the block size is smaller and/or the codebook contains less codevectors. Fig. 2. illustrates the ROI-VQ coding and decoding scheme. Segmentation data
Segmentation ...
Original image
block
Best matching search
index
segmentation information resulting from the image analysis, the encoder and the decoder always uses the appropriate codebook. Thus, the image is compressed with a high quality wherever it is required, and with a low bitrate whenever a higher distortion is permitted.
The proposed compression algorithm has been adapted for echoendoscopic images of the esophagus wall. The Endoscopic Ultrasonography is a very efficient tool in the detection and study of various gastrointestinal cancers. This imaging technique can also be used in the detection of anastigmatic recurrence of tumors, and in the correct evaluation of tumor response to chemotherapy or radiation therapy. The sonographic pattern of the oesophageal wall consists of hypo and hyperechoic layers and there is a good correlation between the echolayers and the hystologic layers of the digestive track wall. The anatomical structure of the oesophageal wall is illustrated in Fig. 3. The actual endosonographic acquisition system displays the ultrasonic response of the tissues on a video monitor. These video images are digitized and stored in a micro-computer. An example of a 2D echo-endoscopic slice is shown in Fig. 4. 1st layer : hyperechogene (Interface) 2nd : Hypoechogene (Mukosa) 3rd : Hyperechogene (Interface)
Region description ... Table look-up Reconstructed image
4th : Hypoechogene (Muscle) 5th : Hyperechogene (Interface)
Fig.3. The structure of the esophagus wall Codebook Codebook ... 1. 2.
Codebook K.
Region codebooks
Codebook Codebook Codebook ... 1. 2. K. Region codebooks
Fig. 2: ROI-VQ encoder and decoder
First of all, an image analysis is necessary in order to determine the regions, to locate the objects. This step may mainly consist of segmentation, object detection (contour detection, classification), etc. Since a strong prior knowledge is exploitable, specified and efficient segmentation techniques can be developed for a given image type. The results of the image analysis is a model, i.e. the global structure of the image. This description is needed to be transmitted to the decoder, and it serves to guide the compression. This data does not take in general a large space although its importance is primary. According to the
Fig.4. An ultrasound esophagus image
Hence, the most important zone of these images is the one corresponding to the esophagus wall, because the diagnosis depends mainly on the content of this region. The surrounding region belongs to several tissues of the human breast, eventually the heart or the aorta. This region can also represent some pathologies, but a slight distortion is tolerable, since only the main structure is important. However, the central part of the image does not contain any relevant information. This zone is black containing several bright circles which do not correspond to any organ but are due to the ultrasound acquisition technique. All these considerations lead to divide the image to be compressed into three parts representing the « empty », the « esophagus » and the « tissues », respectively, as shown in Fig. 5. In [6] there is described a contour detection algorithm specified for esophagus images, which is able to detect the internal contour of the esophagus wall. Thus, using this algorithm the image segmentation is performed: the « empty » is the zone surrounded by the detected contour, the « esophagus » is a fixes width region surrounding the « empty » and the « tissues » are the rest.
Fig.5. ROI-VQ compression scheme for esophagus images
As detailed in the previous section, the result of the segmentation algorithm has to be transmitted to the decoder. In the case of esophagus images this consist of transmitting the detected contour, which allows the decoder to segment the image. The contour can be described as the two coordinates of an arbitrary point (2x9 bits if the image size is inferior to 512), and 3 bits are sufficient to transmit every other adjacent pixel. Since, a given point has only 8 possible neighboring pixels, which can be coded on 3 bits. Experiments showed that the contribution of the contour description to the compressed data amount is very modest, for the studied images it is about 0.02-0.03 bpp. V. Simulation results Ten images were used to create the three separate training sets corresponding to the "empty", the "esophagus" and the "tissues". After detecting the internal contour of the
esophagus wall, the three regions were determined and divided into blocks in order to form the training sets. The Kohonen neural learning method was applied to create every codebook [7]. For the "empty" region 64 codewords of 16x16 pixel size is used resulting in a very high, 1:341.3 local compression rate. Since the esophagus wall contains the most diagnostically relevant information, it must be compressed with a high fidelity. Therefore only 2x2 block size is applied, and the codebook size is 256, leading to a 1:4 local compression rate. In order to obtain a stronger compression, 4x4 block size is proposed for the "tissues". However, the codebook is also larger, containing 1024 codevector, in order to avoid a visually annoying reconstruction quality in this region. Thus, the "tissues" are compressed with a 1:12.8 compression rate. The method was tested on several images lying outside the training set. The results presented in the following concerns the original image shown in Fig.4. Fig. 6 shows the resulting compressed image. The final compression rate is about 1:11. Clearly, the visual image quality is very good on the esophagus wall du to the small applied blocks. Furthermore, the center part of the image is quasi-perfectly reconstructed, which might appear surprising because of the large (16x16) codeword size. However, since this zone is almost the same on every image, the generated "empty" codebook contains perfect patterns. On the "tissues" we can remark a slight blocking artifact, but this distortion is still tolerable (and also can be reduced using some postprocessing method). For comparison purposes we report the JPEG image on Fig. 7. compressed approximately with the same compression rate. The Peak Signal to Noise Ratio (PSNR) values calculated both for the whole images and for the three regions are detailed on Table I. (The PSNR is defined as æ 255 2 ö ÷ PSNR = 10 log10 ç è MSE ø where MSE is the mean square error between the original and the restored image.) As shown, the objective reconstruction quality measure is slightly higher when using ROI-VQ than with JPEG. Furthermore, it is much more important that the critical zone, i.e. the esophagus wall is compressed with a much higher fidelity: ROI-VQ results in an improvement of about +2 dB comparing to JPEG. VI. Conclusions We have presented an image compression method, which deeply exploits the advantages of VQ in medical application environment. Based on prior knowledge and applying the ROI aspect, our approach compresses diagnostically important regions with a very good reconstruction quality. Moreover, a rather high overall
compression rate is obtained due to the strong compression in less important image zones. Since separate VQ codebooks are created to compress different objects, every codebook is well adapted to the given region. Simulation results provided by compressing the studied ultrasound images of the esophagus validated the interests of the proposed approach. Not only a good rate/distortion performance is obtained, but the quality is preserved on the most important part, i.e. on the esophagus wall. Furthermore, the use of separate codebooks permitted to apply such a high block size as 16x16 resulting nevertheless in a very good quality. References [1] A.Gersho, R.M.Gray, "Vector quantization and signal compression", Klower Academic Publisher, Boston, 1992 [2] N.M. Nasrabadi and R.A. King, "Image coding using vector quantization: a review", IEEE Trans.Com., Vol.36, pp. 957-971, Aug. 1988.
[3] J.Vaisey, A. Gersho, " Image Compression with variable Block Size Segmentation", IEEE Transactions on Signal Processing, Voll.40, No 8, August 1992 [4] G.Cazuguel, A.Cziho, B.Solaiman, C.Roux, M.Robaszkiewicz, “Improving Spatial Vector Quantization by use of a Quadtree Scheme. Application to Echoendoscopic Image Compression”, Annual International Conference Of The Engineering In Medicine And Biology Society, pp. 894897, Chicago, USA, 1997 [5] G.Cazuguel, A.Czihó, B.Solaiman, C.Roux, “Medical image compression and analysis using Vector Quantization, the Self-Organizing Map, and the quadtree decomposition” Conference on Information Technology Applications in Biomedicine, Washington, USA, May 1998. [6] F.Pipelier, B.Solaiman, S.Grassin, C.Roux, “A new dynamic contour model: application on ultrasound images”, IEEE Engineering in Medicine and Biology Society, p. 167, 1996.
Fig.6. ROI-VQ compressed image Compression rate: 1 : 10.90 ; PSNR: 31.71 dB
PSNR values [dB] “empty” “esophagus” “tissues” overall
Fig.7. JPEG compressed image Compression rate: 1 : 10.77 ; PSNR: 31.60 dB
ROI-VQ 40.78 31.58 31.14 31.71
JPEG 34.00 29.01 31.76 31.60
Table I. Comparison of ROI-VQ and JPEG
