Nuclear Architecture Analysis of Prostate Cancer via Convolutional Neural Networks

In this paper, we present an approach of convolutional neural networks (CNNs) to identify prostate cancers. Prostate tissue specimen samples were obtained from the tissue microarrays and digitized. For each sample, epithelial nuclear seeds were identified and used to generate a nuclear seed map, i.e., only the location information of epithelial nuclei was utilized. From the nuclear seed maps, CNNs sought to learn the high-level feature representation of nuclear architecture and to detect cancers. Applying data augmentation technique, CNNs were trained on the training data set including 73 benign and 89 cancer samples and validated on the testing data set comprising 217 benign and 274 cancer samples. In detecting cancers, CNNs achieved an AUC of 0.974 (95% CI: 0.961-0.985). In comparison with the approaches of utilizing hand-crafted nuclear architecture features and the state of the art deep learning networks with standard machine learning methods, CNNs were significantly superior to them (p-value <; 5e-2). Moreover, stromal nuclei were incapable of improving the cancer detection performance. The experimental results suggest that our approach offers the ability to aid in improving prostate cancer pathology.