CONTEXT BASED MEDICAL IMAGE COMPRESSION WITH TRANSFORM CODING

Abstract

The medical image data acquired from various modalities, such as magnetic resonance imaging (MRI), computed tomography (CT), ultrasonography (US), X-ray imaging (XR) etc. comprise huge volume rendering it impractical for storage and transmission. Even in an average sized hospital, terabytes of digital image data is generated every day and almost all of this data have to be archived. Further, because of the large volume of the medical data, the transmission problem is aggravated in wide area network (WAN) applications which often include low bandwidth (BW) channels, such as long distance telephone lines or an integrated services digital network (ISDN). Therefore, the compression of the medical image data becomes an important area of concern for digitization and reduction of the volume of medical image information and handling of this bulk data generated everyday. In addition, the compression of the medical image data can also facilitate the new emerging trends in medical science such as telemedicine, teleradiology, teleconsultation and the digitized medical information systems (DMIS). Among the available image compression techniques, one of the major problems arises in deciding the compression strategy for a particular application. Hence, the compression decision is largely a balance between archival costs and image quality along with the computational complexity and the rate-distortion performance. Therefore, the choice of an efficient medical image compression (MIC) scheme is a complex trade off the system andthe clinical requirements as well as compression rates (CR) and the image quality. For a high quality diagnostic application, no loss in the reconstructed image is desired, but this will have a very low CR of 3:1 which is inappropriate for bulk medical image data storage and transmission as well as for the picture archiving and communication system (PACS). This type of compression comesunder the category of'lossless compression'. Onthe otherhand, if CR is very high, then there is a possibility of losing some useful information which may be important for the diagnosis and is always a risky task for the medical use because of the importance of the information associated with the medical image and its diagnostic importance. This aspect of compression comes under the category of Tossy compression'. In the conventional compression methods, the entire medical image is compressed at equal level of compression to the diagnostically useful area called region of interest (ROI) as well as rest of the image called background (BG). As a result, the reconstructed image is of xvn poor visual quality at higher CRs. However, in MIC, it is desired to preserve a high quality of ROI as compared to the BG which is generally not available in the conventional coding methods such as the joint photographic experts group (JPEG) based on the discrete cosine transform (DCT) as well as embedded zerotree wavelet (EZW), set partitioning in hierarchical trees (SPIHT) and the set partitioning embedded block (SPECK) coding based on the wavelet transform. The main drawbacks of the conventional compression methods based on the DCT are - blocking and blurring artifacts in the reconstructed image, inefficiency of the DCT to convert the image parameters into fewer coefficients and subdivision of the image into subblocks (8x8 or 16x 16 etc.) causing the distortion at higher CRs. Because of these facts, the compression methods based on the DCT have poor rate-distortion, rate-correlation and low qualityperformance of the compressed medical image data at higher CRs as compared to their counter parts based on the wavelet transform (WT). One of the first methods of ROI coding is incorporated in JPEG2000 (JP2K) which is an extension of the JPEG. Basically, there are two ROI coding methods supported by JP2K namely the maxshift method (MSM) and general scaling method (GSM) which are based on embedded block coding with optimized truncation (EBCOT). However, these methods are limited in application and have an approximation of ROIs as well as do not support exact and multiple ROIs. TheMSMapproach exploits the possibility of representing an ROI by shifting its coefficients on the top of the BG coefficients which produces a separate ROI-BG bitstream. The GSMrealizes a coefficient ranking by using a sub-band coefficient weighing mask prior to the quantization stage, producing a bitstream where the ROI and BG are coded together with a different quality level. However, these methods have several drawbacks such as rough ROI selection in place of exact ROIs, no multiple ROIs, lack of the flexibility to allow an arbitrary scaling value to define the relative importance of the ROI and BG coefficients. In addition, different ROIs can not have their own scaling values as well as the bitstream overflow problem in the MSM. Further, the GSM requires the generation of an ROI mask and distinction of ROI-BG coefficients at both encoder and decoder sides which increases decoder's complexity and processing overhead. In the view of above limitations, the present work is focused on the selective ROI coding known as 'context based medical image compression' (CBMIC) which addresses the issue of bulk storage and faster data transmission requirements of the medical image information while maintaining an acceptable diagnostic image quality at high CRs. The other important reason for selection of the CBMIC is high spatial resolution and contrast sensitivity requirements ofthe medical image data analysis. Because of the importance of medical image xvin information from diagnostic point of view, the care has been taken while compressing that it does not lose any of the important features those may be required for the diagnosis of an image by a medical expert. These features have been assured by selecting efficient contextual region of interest (CROI) compression strategy which targets overall higher CRs and better reconstructed diagnostic image quality. The diagnostically useful region of a medical image is defined as the 'CROI' in this work. The proper selection of the CROI extraction methodology and the bit rate (bpp) allocation in the CBMIC can give better quality of image reconstruction along with high CRs. In practice, the better is CROI selection methodology; the higher will be accuracy of compression and easier will be to diagnose the image by the medical experts. The proposed CBMIC methods here are mainly based on the WT and DCT in which the medical image is subdivided into the useful regions from the diagnosis point of view i.e. 'CROI' and the reaming part of the image as 'BG'. These regions (CROI and BG) are separated by suitable proposed algorithms based on binary mask generation, morphological, segmentationand interactive method. After extraction of the CROI coefficients, the remaining BG coefficients are compressed at very high CR while the CROI coefficients are compressed at very low CR to avoid loss of any useful information along with getting high overall CRs of the compressed image data. As a result, the reconstructed medical image after compression retains all the important features desired for the diagnostic purpose with high compression efficiency (CE) and better image quality. Thereby, the proposed CBMIC techniques meet the required features of an efficient compression methodology i.e. good reconstructed image quality with high CRs, low mean square error (MSE), high peak signal to noise ratio (PSNR), very high correlation coefficients (CoC), high mean opinion scores (MOS) and low percentage rate distortion (PRD). In the present work, the MOS of reconstructed images, has been evaluated by six different radiologists and the medical experts consulted for different case studies undertaken on a scale of 0 to 10. In the present work, four CBMIC algorithms are proposed namely 'context based discrete cosine transform' (CBDCT) coding, 'context based discrete wavelet transform' (CBDWT) coding, 'contextual set partitioning in hierarchical trees' (CSPIHT) coding and 'contextual JPEG2000' (CJP2K) coding which have been implemented on four different types of radiological image modalities (viz. US, MR, CT and XR) and the standard available natural images (Lena, Barbara, cameraman etc.) each consisting of twenty five case studies. In further part of the compression, the non-context based coding methods namely JPEG based on DCT, discrete wavelet transform (DWT) based embedded zerotree wavelet (EZW), set partitioning in hierarchical trees (SPIHT) and the JP2K coding methods have been xix implemented on the same images for comparison of the performance of the proposed algorithms. The comparative analyses of the above proposed CBMIC techniques have also been carried out to find the most suitable algorithm for a specific medical image modality. The proposed CBMIC algorithms have been implemented by using image processing and wavelet toolboxes of the MathWorks in MATLAB (R2008b). The success of any compression algorithm depends upon the output image quality metrics (IQM) performance parameters and the retrieved image fidelity. Therefore, an exhaustive performance analysis of the obtained results for the proposed context based coding algorithms (viz. CBDCT, CBDWT, CSPIHT and CJP2K) with the existing non-context based standard coding methods (viz. JPEG, EZW, SPIHT and JP2K) have been carried out in the present work on the basis of standard subjective and objective IQM performance parameters namely CR, bpp, MSE, PSNR, CoC, PRD and the reconstructed image quality based on the MOS for the CRs ranging from 8:1 to 128:1 (bpp from 1.0 to 0.0625) for all the image modalities considered. The histogram matching is also performed to correlate the pixel densities and the related gray levels of the original and the reconstructed images. Also, the graphical and the reconstructed image comparisons have been carried out for the better analysis of the results. The IQM performance parameters of the proposed CBMIC algorithms are also compared with the other existing standard ROI coding methods (viz. EBCOT, GSM, MSM and Implicit) for the standard images. An overview of the results obtained from these proposed methods is presented in the following paragraphs. The first proposed CBMIC algorithm i.e. CBDCTis implemented by using binary mask generation of CROI extraction method and 8x8 block size selection and has been tested on various image modalities considered as mentioned above. The over all bit rate variation is taken from 1.0 to 0.0625. However, the CRs for the CROI is restricted from 4:1 to 16:1 whereas it is extendedup to 256:1 for the BG to get an over all satisfactory CR in the range of 8:1 to 128:1 as well as the diagnostically acceptable image quality judged on the basis of MOS and IQM performance parameters. For the proposed CBDCT coding, the improvements in the results for few cases in each image modality are given as follows: For the proposed CBDCTalgorithm at common bpp=0.0625, in case of US1 image, the improvement in MSE, PSNR, CoC and MOS is 74.3329 (21.89%), 1.0734 dB (4.70 %), 0.042786 (4.57 %) and 2.4500 (40.27%), respectively as compared to the non-context based DCT compression. For another imagemodality e.g.MR3 image the improvement in the MSE, PSNR, CoC and MOS at the same bpp is 20.4321 (17.16 %), 0.8178 dB (2.99 %), 0.052065 (5.56 %) and 2.8334 (48.57 %), respectively. Similarly, for CTl image, the improvement in xx the MSE, PSNR, CoC and MOS is obtained as 113.1487 (73.47 %), 5.7594 dB (21.98 %), 0.051282 (5.43 %,) and 2.35 (38.11 %) respectively. In case of the fourth image modality i.e. XR2 image, the improvement in the MSE, PSNR, CoC and MOS is obtained as 34.5893 (43.29%), 2.4630 dB (8.46 %), 0.071024 (7.66 %) and 2.8083 (48.84%), respectively. Similarly, in case of the standard Lena image, the improvement in the MSE, PSNR, CoC and MOS is obtained as 67.2580 (68.42%), 5.0054dB (17.75%), 0.064592(6.94%) and 2.1250 (33.55 %), respectively. From these reported results, it is found that there is significant improvement in the IQM performance (both the objective and subjective) and image quality of the proposed CBDCT algorithm as compared to the conventional non-contextual DCT based JPEG coding at low bit rates (below 1.00) for all the image modalities considered. Although, the case of only one image has been given here for each modality for illustration, but the similar trend is observed on many more images considered under test used for such purpose of consisting of all the above image modalities. The important feature of the proposed CBDCT algorithm is that as the bit rate reduces (from 1.00 to 0.0625) the decline in the image quality and the IQM performance is slow in the CBDCT as compared to the JPEG compression. In addition, the blocking and blurring artifacts occurring at higher CRs in the reconstructed image in case of the JPEG compression have also reduced significantly in the proposed CBDCT coding, as the CROI is compressed very lightly (4:1<CR< 16:1). The second proposed CBMIC algorithm is CBDWT which is based on the contextual embedded zerotree wavelet transform coding and the morphological CROI mask generation method. In the CBDWT coding, three-level decomposition algorithm is used which is based on the biorthogonal wavelet family (biorl.l to bior6.8). The wavelet coefficients are encoded progressively on the priority basis i.e. the CROI coefficients get higher priority over the BG. The important feature of this method is that the CROI coefficients are quantized at very low thresholds while the BG coefficients are quantized at higher thresholds to get over all good CRs as well as maintaining better reconstructed (diagnostic) image quality. In this case also, all the same image modalities are considered as in the above case for the testing of the proposed algorithm. For the CBDWT, the improvements in the results for few cases in each image modality as comparedto the non-contextual DWTcoding are reported as follows: For the proposed CBDWT algorithm at bpp=0.0625, in case of US1 image, the improvement in MSE, PSNR, CoC and MOS is 61.4978 (49.72%), 2.9835 dB (11.05 %), 0.042786 (0.0484 %) and 2.2417 (35.40%), respectively as compared to non-context based DWTcompression. For another image modality i.e. MR3 image the improvement in the MSE, xxi PSNR, CoC and MOS at the same bpp are 198.8675 (91.28 %), 10.5922 dB (42.81 %), 0.009394 (0.9634 %) and 2.2666 (35.32 %), respectively. Similarly, for CTl image, the improvement in the MSE, PSNR, CoC and MOS are obtained as 102.3290 (86.72 %), 8.5581dB (31.12 %), 0.011237 (1.1368 %,) and 2.125 (33.01%), respectively. In case of the fourth image modality i.e. XR2 image, the improvement in the MSE, PSNR, CoC and MOS is obtained 83.2170 (89.48%), 29.6216 dB (33.88%), 0.002731 (0.2739 %) and 2.5084 (41.23%), respectively. Similarly, in case of Lena image, the improvement in the MSE, PSNR, CoC and MOS is obtained as 186.7027 (75.93%), 6.1853 dB (25.54 %), 0.004312 (0.4342 %) and 2.0250 (31.07%), respectively. In the proposed CBDWT coding, it is observed that the IQM performance is excellent along with good reconstructed image quality at all the bit rates. Further, as the bpp reduces from 1.00 to 0.0625, there is very low decline in the image quality and the IQM performance of the CBDWT as compared to the non-contextual DWT coding. The proposed algorithm performs equally well for all the image modalities considered. Here, again for illustration the results of only one case in each modality has been presented, but the similar trend has been observed for all the images considered. Also, the proposed method gives high PSNRs and MOSs at low bit rates indicating the good improvements in the image quality which is a major draw back of the conventional methods (JPEG, EZW). In the proposed CBDWT method, the blocking and blurring artifacts are also eliminated as there is no need of tiling of the image into subblocks (8x8 or 16x16 etc.). The same methodology of compression has been adapted for the third and the fourth proposed CBMIC algorithms namely CSPIHT and the CJP2K and the similar improvements in the results are obtained. The CSPIHT coding works on the principle of the contextual set partitioning in the hierarchical tress and the segmentation method of CROI extraction by using Sobel's and Canny's edge detectors. The 'Symlet' (sym4 to sym8) and Daubechies (db4 to dblO) families of wavelets have been used for the analysis and synthesis part of the transform based contextual coding. From the exhaustive result analysis of the context and non-context based coding techniques (SPIHT and CSPIHT), it is found that the IQM performance in the proposed method is significantly improved at all the bit rates ranging from 1.00 to 0.0625. It is also observed that as the bit rate reduces from 1.00 to 0.0625, there is fast deterioration in the image quality and the IQM performance in case of the SPIHT whereas this decline is significantly reduced and it is consistent in case of the CSPIHT. The proposed CSPIHT method overcomes the limitations of the conventional methods (SPIHT, EZW and JPEG) so far as the IQM performance is concerned. xxn The fourth proposed CJP2K algorithm is based on the embedded block coding and utilizes the interactive method of CROI extraction with the help of a graphical user interface. Here, the biorthogonal (biorl.l to 6.8) and the Daubechies 9/7 families of the wavelets have been used. In this case also, the CROI and BG regions are compressed with different quantization thresholds i.e. BG high and that of CROI low. As a result, the overall good CRs are achieved along with improved image quality. In the conventional methods (JP2K, SPIHT, EZW and JPEG), as the bpp reduces, the image quality declines fast. However, in the CJP2K, the decline is not as much and there is significant improvement in the IQM performance parameters in the bpp range from 1.00 to 0.0625 with a better quality of the reconstructed image. Among the implemented coding algorithms (four in contextual mode and four in noncontextual mode), it is found that the context based techniques (viz. CBDCT, CBDWT, CSPIHT and CJP2K) perform better for all the image modalities considered than that of noncontext based (viz. JPEG, EZW, SPIHT and the JPEG2000) coding techniques for the bit rate from 1.00 to 0.0625. Also, among the proposed context based methods (or contextual methods), it has been observed that the proposed CBDWT, CSPIHT and CJP2K algorithms perform better at desired lower bit rates (below 0.125) whereas the CBDCT algorithm performs better at moderate bit rates (above 0.25). The CBDWT and CSPIHT algorithms have outperformed the CBDCT, CJP2K algorithms in terms of better IQM performance parameters and the visual quality of the reconstructed images for the same bpps. The CSPIHT has given excellent improvement in PSNR, MOS and CoC, and reduction in MSE for the acceptable image quality and high CR values which is better than the other three methods. Therefore, in comparison to four proposed context based coding methods, the CSPIHT gives the best IQM performance and the image quality whereas the CBDWT gives second best performance followed by the CJP2K and the CBDCT algorithms. The proposed methods perform excellently well on both the natural as well as on medical (radiological) images. However, the IQM performance and image quality varies from image to image and from coder to coder. The improvements observed have been better in those image modalities which have larger BGs than the CROIs. The only limitation of the proposed techniques is that there is a possibility of the blurring at the boundaries of the BG and the CROI regions but it does not affect the image quality as blurring region will be out of the CROI. As a conclusion, the prime contribution of the present research work has been the suggestions of the modified CBMIC methods which provide improvement in the IQM performance and a very high diagnostic quality of the reconstructed image with reduced xxm blocking and the blurring artifacts and hence can reduce heavily transmission and the storage costs of the huge medical data generated every day and will be well suited for the limited bandwidth networks. Thus, the proposed methods will prove to be very useful in high quality diagnostic requirements of medical image data without losing any useful information and have promising application in the emerging fields of medical science such as teleradiology, telemedicine, teleconsultation, e-health, digitized medical information systems, PACSs and the statistical analysis of medical image information for the future predictions as well.