LEARNING BASED SUPER-RESOLUTION OF HISTOLOGICAL IMAGES

For better perception and analysis of images, good quality and high resolution (HR) are always preferred over degraded and low resolution (LR) images. Getting HR images can be cost and time prohibitive. Super resolution (SR) techniques can be an affordable alternative for small zoom factors. In medical imaging, specifically in the case of histological images, estimating an HR image from an LR one requires preservation of complex textures and edges defining various biological features (nuclei, cytoplasm etc.). This challenge is further aggravated by the scale variance of histological images that are taken of a flat biopsy slide instead of a 3D world. We propose an algorithm for SR of histological images that learns a mapping from zero-phase component analysis (ZCA)-whitened LR patches to ZCA-whitened HR patches at the desired scale. ZCA-whitening exploits the redundancy in data and enhances the texture and edges energies to better learn the desired LR to HR mapping, which we learn using a neural network. The qualitative and quantitative validation shows that improvements in HR estimation by proposed algorithm are statistically significant over benchmark learning-based SR algorithms.