PhD seminar, Ida Arvidsson
Plats:Sigma, MH:210b (library)
Title: Prostate Cancer Classification using Deep Learning
Abstract: Prostate cancer is the most common cancer diagnosis among men. Today the diagnosis is determined by pathologists based on ocular inspection of stained slices of prostate biopsies in a light microscope. To aid pathologists in analyzing the samples, and to make their diagnoses come closer to consensus, we are developing methods to automatically detect and classify the malignant areas. In this talk I will present a convolutional neural network which we have designed and trained from scratch for the task of Gleason grading, i.e. classification of the malignant tissue depending on the severity of the cancer. This has turned out to be very successful for images from the same hospital as the dataset used when training the network, while the results are very poor when the network is applied to other datasets due to e.g. inevitable stain variations. To improve the performance on unseen datasets we have investigated some different approaches, for example digital stain separation and normalization based on singular value decomposition. I will present our results this far, and discuss our current ideas.