Visually lossless compression of volumetric medical images

Abstract

Image compression techniques are broadly classified into two categories. One is lossless or reversible compression and another is lossy or irreversible compression. Lossless or reversible compression techniques are more preferred by medical professionals as every detail that can be perceived by a medical expert is very significant in the diagnosis of any pathological condition. However, lossless methods for compression of medical images don’t promise good compression efficiency. On the other hand, a number of lossy methods have been proposed that provide high compression, but at the cost of quality degradation, which is unacceptable by the medical experts. In summary, medical image compression demands for high visual quality, close to that of lossless methods and good compression efficiency close to that of lossy methods. To achieve high visual quality with good compression performance, it is very important to utilize the Human Visual System (HVS) characteristics. Exploiting HVS characteristics guarantees the removal of visually insignificant data without altering any diagnostically significant data. Most of the medical images exhibit bilateral symmetry inherent in the human body. Additional compression can be achieved by exploiting the symmetry if present in images. Further compression can be achieved by means of Volume of Interest (VOI) based hybrid coding. In this method diagnostically important regions in all slices are losslessly coded while the rest of the image is coded using lossy methods. This could be beneficial in the area of tele-medicine with bandwidth limitations. This thesis proposes five visually lossless compression methods (VLIC-1 to VLIC- 5) for 3-D Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images that embeds Just Noticeable Distortion (JND) profile to remove visually redundant information. The first three algorithms are developed in spatial domain in which background luminance and texture based JND models are used to eliminate perceptually redundant information. The first algorithm is based on predictor and block matching technique, while second and third algorithms are based on block matching technique. Symmetry nature of human anatomy reflected in medical images is exploited in all the three algorithms. The fourth algorithm is developed in transform domain using wavelets. Along with this, wavelet based vision model is utilized to remove visual redundancies. Final algorithm is a hybrid VOI based method, where in wavelet based visually lossless method is used in diagnostically important regions/volumes and Discrete Cosine Transform (DCT) based lossy technique is used in other regions. Algorithms are tested on several sets of MRI, CT, x-ray angio images obtained from various sources. Results obtained with 3157 slices of CT images, 886 slices of MR images, 608 slices of x-ray angio slices and 587 slices of MRI brain images with tumor are documented to compare the performance. Compression Ratio (CR) and quality of the encoded images are compared with lossless and lossy algorithms.FEBRAURY HVS based quality metrics are also computed for the above algorithms. The obtained results show better bit per pixel against lossless compression techniques, without any degradation in visual quality. In case of Differential Pulse Code Modulation (DPCM) and symmetry based visually lossless compression technique (VLIC-1), average % reduction in bit per pixel is 11.8 compared to 3D context based Embedded coding of Zerotrees of Wavelet coefficients method (Bilgin et al. 1998). In case of symmetry based visually lossless compression technique (VLIC-3) average % reduction in bit per pixel is 6.29 compared to medical image lossless compression method (Ait Aoudia et al. 2006). Wavelet based visually lossless coder (VLIC-4) reduces the bit rate by 20.66% compared with DPCM, 28.24% compared with high efficiency video coding (Sanchez and Bartrina-Rapesta 2014), 14.60% compared with Joint Photographic Experts Group (JPEG)-2K, 13.28% compared with JPEG-LS, and 46.85% compared with JPEG-3D (Bruylants et al. 2015). The maximum variation in Structural SIMilarity index value for reconstructed images is from 1 to 0.99, indicating they are visually lossless. CR is found to be within the acceptable compression range as suggested by Royal College of Radiologists and European Society of Radiology. The subjective quality of the VOI scheme is validated by radiologist.