Super Resolution of Medical Image Using Integrated Approach of Deep Learning and Sparse Coding

Abstract

Image Super-Resolution is the active field of image processing. Images and videos with high resolution are used in various fields like medicine, agriculture, pattern recognisation etc for better analysis. The need of enhancement of the resolution of image is motivated due to the advancement of the pictorial information used by humans for interpretations, for autoarchic machine applications and for the purpose of efficient storage and transmission. One of the major fields is medical imaging. As medical images are more sensitive to noise, so to enhance the intensity of the image is turn out to be a major area. The High resolution medical images can localize any disease or analyzes the body part more accurately. There are many applications for increasing the resolution of image but they are not very effective as they add physical artifacts such as noise and blur. There are various algorithms of image super resolution which uses dictionaries for reconstruction and these dictionaries have to train explicitly. This thesis proposes a method to increase the resolution of a medical image. In this framework conventional sparse coding model has been extended with key concepts of deep learning. This framework doesnot require training of dictionaries explicitly. Firstly, a low resolution medical image is denoised to remove noise. As medical images are more sensitive to noise due to the acquisition devices. Then this denoised image is inputted to the network where features are extracted for each LR patch. Then LR patch is fed into a network. This network is based on learned iterative shrinkage and thresholding algorithm (LISTA) whose layers strictly correspond to each step in the processing flow of sparse coding based image SR. This way sparse representation technique is effectively encoded in our network structure, and at the same time all the components of sparse coding can be trained jointly through back-propagation. In the next layer sparse code obtained is multiplied with dictionary to reconstructs HR patch and then the recovered patches are placed back to their respective positions in the HR image by the fixed layer which aligns pixels in overlapping patches. At last, a HR medical image is obtained as output. The performance evaluation of the proposed method is based on Peek Signal to Noise ratio (PSNR), Structure Similarity Index (SSIM) values. The PSNR and SSIM values of our proposed algorithm are better than the Bicubic interpolation, Sparse coding algorithm for both noiseless and noisy image etc. The average PSNR value of proposed method is more than that of sparse coding by 0.27 for noiseless image.