2019/01/18 by Francis Tom, Himanshu Sharma, Tom, Francis +7 · 1 citation
Biochemistry, Genetics and Molecular Biology · Computer Science · Engineering · #Advanced Image Processing Techniques #Cell Image Analysis Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Image Processing Techniques and Applications
paper · pdf · doi:10.48550/arxiv.1901.06405
openalex publication_date 2019/01/18 · openalex created_date 2022/07/30 · openalex updated_date 2026/07/28
Adversarially trained deep neural networks have significantly improved\nperformance of single image super resolution, by hallucinating photorealistic\nlocal textures, thereby greatly reducing the perception difference between a\nreal high resolution image and its super resolved (SR) counterpart. However,\napplication to medical imaging requires preservation of diagnostically relevant\nfeatures while refraining from introducing any diagnostically confusing\nartifacts. We propose using a deep convolutional super resolution network\n(SRNet) trained for (i) minimising reconstruction loss between the real and SR\nimages, and (ii) maximally confusing learned relativistic visual Turing test\n(rVTT) networks to discriminate between (a) pair of real and SR images (T1) and\n(b) pair of patches in real and SR selected from region of interest (T2). The\nadversarial loss of T1 and T2 while backpropagated through SRNet helps it learn\nto reconstruct pathorealism in the regions of interest such as white blood\ncells (WBC) in peripheral blood smears or epithelial cells in histopathology of\ncancerous biopsy tissues, which are experimentally demonstrated here.\nExperiments performed for measuring signal distortion loss using peak signal to\nnoise ratio (pSNR) and structural similarity (SSIM) with variation of SR scale\nfactors, impact of rVTT adversarial losses, and impact on reporting using SR on\na commercially available artificial intelligence (AI) digital pathology system\nsubstantiate our claims.\n