IMAGE COMPRESSION FOR TELEMEDICINE

Abstract

Telemedicine is a new practice of medicine over distance. One of the important aspects of telemedicine is medical image transmission between and among healthcare organizations. Most of the medical image modalities such as, CT, MR, PET etc. occupy large amount of binary data size as multiple slices are generated in single examination. Therefore, efficient image compression techniques are inevitable for telemedicine applications. Lossless compression technique is generally preferred for medical images as it prevents the possibility of losing diagnostic information by preserving every relevant detail of the image. However, the compression ratio obtained by lossless image compression techniques are very less. Near lossless compression is a lossy compression technique, in which the maximum peak absolute error (PAE) of the reconstructed image can be set to a desired specific value. It produces higher compression ratio than the lossless technique, with the control over the maximum absolute error of each pixel in the image. Region of Interest (RoI) coding technique, takes the advantage of both lossy and lossless techniques. In this method, diagnostically important areas called Region of Interest are compressed in lossless way or with high quality, whereas the unimportant areas such as background are compressed in lossy way. The RoI coding achieves high compression ratio without loss of diagnostically important information. Resolution scalable image coding is useful in applications such as telebrowsing when several images of low resolution are displayed together for comparison at first and later magnified into a higher resolution. The image with a reduced resolution is transmitted at first, then the information required to obtain a higher resolution of the same image is transmitted progressively. The present work is aimed at achieving four objectives. The first objective has been to develop a lossless image compression algorithm which can provide higher compression ratio than the state-of-the-art techniques. Lossless image compression is most generally performed in two stages. In the first stage, image data is decorrelated and in the second stage, the decorrelated data is entropy coded. The entropy coding algorithms developed over the past years such as arithmetic coding, have achieved compression performance very close to its theoretical bound. Thus, the research objective is more focused at the image data decorrelation stage. A novel dual level DPCM namely DL-DPCM is proposed for efficient decorrelation of ii image data. The DL-DPCM consists of a linear DPCM cascaded by a neural network based non-linear DPCM namely, context adaptive switching neural network predictor (CAS-NNP). A lossless compression technique namely ‘DLD-LS’ was developed using the DLDPCM. In this scheme, the image is encoded sequentially, pixel by pixel, in a raster scan order. The inter-pixel redundancies present in the input image data are removed in two stages by two DPCMs which are cascaded. The ‘2D-LDPCM’ produces a prediction error image, which is the difference between the original image and the predicted image based on a subset of previously encoded pixels called the causal template. The CAS-NNP, which is adaptive and nonlinear, further removes the redundancies present in the error image. In the predictor stage, the pixels in an image are divided into plain, gradient and edge regions. Three different NN predictors are used so that each predictor is optimized to predict pixels in different regions for achieving higher overall prediction accuracy. The prediction error image of CAS-NNP is entropy coded using context adaptive arithmetic coding after pre-processing to produce a binary output codestream. Comparative analysis is done with standard lossless image compression techniques, ‘CALIC’ and ‘LOCO-I’. The developed algorithm was tested on five different datasets consisting of CT, MR, PET and Angiogram. It is observed from the experiments that the proposed DL-DPCM has provided an improvement in the first order entropy values of the prediction error, which has resulted in lower bits per pixel (bpp) value of the lossless encoder. From the comparative analysis with predictors, ‘GAP’ (used in CALIC) and ‘MED’ (used in LOCO-I), it is observed that DLDPCM achieved an average improvement of 0.34 bpp and 1.20 bpp in first order entropy values for CT images, which are 12.9% and 32.7%, respectively. Similarly, for MR images the improvements were 0.64 bpp and 1.88 bpp, which are 17.7% and 38.2% improvement compared to GAP and MED. For PET images the improvements were 0.52 bpp and 0.97 bpp, which are 34.8% and 49.6% improvement compared to GAP and MED. Similarly, for angiogram sequences the improvements were 0.04 bpp and 1.40 bpp, which are 1.43% and 33.3% improvement compared to GAP and MED. The bit rate per pixel improvement after entropy coding the prediction error were 0.07 bpp (4.3%) compared to CALIC and 0.16 bpp (9.0%) compared to LOCO-I for CT images. Similarly for MR images, the improvements were 0.21 bpp and 0.41 bpp which are 7.4% and 13.4% improvement compared to CALIC and LOCO-I. For PET images, the improvements were 0.17 bpp and 0.27 bpp which are 23.4% and 31.7% improvement compared to CALIC and LOCO-I. Similarly for angiogram sequences, the performance was slightly less than iii CALIC with 0.01 bpp (0.5%) extra requirement but showed 0.06 bpp (2.6%) improvement over LOCO-I. In the second stage, the DLD-LS was extended to perform near lossless compression. This is achieved by including a quantization stage after the 2D-LDPCM. The prediction error image produced by the 2D-LDPCM is quantized depending on the quality of the reconstructed image required. The developed algorithm was tested on CT, MR, PET and angiogram. The performance of the developed algorithm was compared with the near lossless compression performance of CALIC and LOCO-I in terms of PSNR values. The overall bit rate per pixel performance of the developed near lossless compression algorithm was better than CALIC, LOCO-I, JPEG and JPEG2000, for same maximum PAE. Further, for low PAE values, the developed near lossless compression algorithm showed better performance in terms of the parameters, PSNR, MSSIM and UQI. In the third stage, the DLD-NLS was extended to perform region of interest coding. The region of interest were compressed without loss of information and the area other than the region of interest were compressed with near lossless compression technique. The algorithm is realized by quantizing the prediction error image only on the selected areas defined by an RoI mask. Region of interests with 5%, 10% and 25% of the image area with square shape at the middle were defined. The CT, MR, PET and angiogram were used to test the RoI coding capabilities of the developed algorithm. It was observed that by using the developed RoI coding, bit savings of 0.75 bpp, 0.61 bpp and 0.53 bpp were achieved when 5%, 10% and 25% of the image were marked as RoI respectively and PAE = 1 was allowed for background compared to lossless compression without RoI coding. In the fourth and final stage, a resolution scalable image compression algorithm was developed using the DL-DPCM. To achieve resolution scalability, firstly the image was scaled down to the desired level by sub-sampling, which represents the lowest level of resolution. The image at the lowest resolution level is coded first using 2D-DPCM and context adaptive arithmetic coding. Similarly, the image at the next higher level is encoded. This procedure is followed progressively till the original resolution of the image. Near lossless compression and region of interest coding capabilities were incorporated by adding a quantization stage after the DPCM stage. The proposed algorithm, ‘RDLD’ was designed to perform lossless compression (RDLD-LS), near lossless compression (RDLD-NLS) and region of interest coding (RDLDRoI). The compression performance of the RDLD-LS, RDLD-NLS and RDLD-RoI were compared with benchmark compression algorithms by testing on CT, MR, PET and angiogram. iv The overall compression performance of the developed resolution scalable coding algorithm was better than JPEG2000 in lossless mode. An improvement of 0.09 bpp (4.05%) was achieved for CT images and a very slight improvement of 0.002 bpp (0.42%) was achieved for MR images. The average improvement on the total data set was 0.03, which is 1.47% improvement over JPEG2000. Further, the developed algorithm achieved lower maximum PAE compared to JPEG2000 for different compression ratios.