Translational Technologies

Computational pathology

 

Tissue diagnosis is a mix of science and art because disease biology is heterogeneous and human visual interpretation is subjective. As a result, the diagnostic pipeline is slow, error-prone and highly subjective. To ease the pathologist’s cognitive burden and minimize the human error, I have been working on two ongoing projects in visual learning and tissue pattern recognition from histopathology images: (i) learning the origins of a tissue (e.g., breast, kidney, lung, etc.) from its spatial architecture; and (ii) segmenting and triaging regions of interest for diagnosing atypical ductal hyperplasia in breast cancer. In the first project, we experiment with deep neural networks for classifying tissue origins and show that its performance is comparable to that of expert pathologists. In the second project, we use both deep and shallow learning to classify these histological structures for ranking of cancer risks. The deep network achieves testing accuracy of 72.5% on a 4-way classification task (normal and stages of atypia ducts) and the shallow learning algorithm achieves testing accuracy of 80% on a 2-way classification task (normal vs. atypia ducts).