2019/10/14 by Bairavi Venkatesh, Venkatesh, Bairavi, Tosha Shah +5
Computer Science · Biochemistry, Genetics and Molecular Biology · #AI in cancer detection #Generative Adversarial Networks and Image Synthesis #Cell Image Analysis Techniques
paper · pdf · doi:10.48550/arxiv.1910.06428
It is common for pathologists to annotate specific regions of the tissue,\nsuch as tumor, directly on the glass slide with markers. Although this practice\nwas helpful prior to the advent of histology whole slide digitization, it often\noccludes important details which are increasingly relevant to immuno-oncology\ndue to recent advancements in digital pathology imaging techniques. The current\nwork uses a generative adversarial network with cycle loss to remove these\nannotations while still maintaining the underlying structure of the tissue by\nsolving an image-to-image translation problem. We train our network on up to\n300 whole slide images with marker inks and show that 70% of the corrected\nimage patches are indistinguishable from originally uncontaminated image tissue\nto a human expert. This portion increases 97% when we replace the human expert\nwith a deep residual network. We demonstrated the fidelity of the method to the\noriginal image by calculating the correlation between image gradient\nmagnitudes. We observed a revival of up to 94,000 nuclei per slide in our\ndataset, the majority of which were located on tissue border.\n