Prostate cancer diagnosis by deep learning
Shuijing Liu · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2018
Prostate cancer diagnosis by biopsy images of human tissue requires experienced trained pathologists and the cost is high. To facilitate prostate cancer diagnosis, we built and trained binary classifiers using deep convolutional neural networks (CNNs) on two datasets: one contains cancerous and healthy biopsy images of prostate tissues (referred as Dataset1), and the other contains biopsy images of tissues with recurred cancer and fully recovered tissues (referred as Dataset2). We extracted patches from biopsy images of human tissues, then built and trained CNN models to classify the patches. We achieved 82% test accuracy on Dataset1 and 63% accuracy on Dataset2. In addition, we used ensemble methods to further boost the performance. With predictions of all patches in our datasets, we performed majority voting on the image level, and the accuracy increases by 5% to 10% on the first dataset. Then we used Bootstrap Aggregation (Bagging) to further increase accuracy to 100% on Dataset1. However, the two-step ensemble methods above have little influence on the accuracy of Dataset2. When visualizing the predictions on the second dataset returned by our models, no clear patterns are found that can distinguish the two classes.