Automated Gleason Grading and Gleason Pattern Region Segmentation Based on Deep Learning for Pathological Images of Prostate Cancer

Prostate cancer is the second-deadliest cancer in men in the United States, seriously affecting people's life and health. The Gleason grading system is one of the most reliable methods to quantify the invasiveness of prostate cancer, which is of great significance for risk assessment and treatment planning for patients. However, the task of automating Gleason grading is difficult because of the complexity of pathological images of prostate cancer. This paper presents an automated Gleason grading and Gleason pattern region segmentation method based on deep learning for pathological images of prostate cancer. An architecture combining the atrous spatial pyramid pooling and the multiscale standard convolution is proposed for the segmentation of the Gleason pattern region to get accurate Gleason grading. In addition, the postprocessing procedure based on conditional random fields is applied to the prediction. The quantitative experiments on 1211 prostate cancer tissue microarrays demonstrate that our results have a high correlation with the manual segmentations. The mean intersection over union and the overall pixel accuracy for the Gleason pattern region are 77.29% and 89.51%, respectively. Furthermore, the results of the automatic Gleason grading were comparable to the results of experienced pathologists. The inter-annotator agreements between the model and the pathologists, quantified via Cohen's quadratic kappa statistic, was 0.77 on average. Our study shows that the method of combining different deep neural network architectures is suitable for more objective and reproducible Gleason grading of prostate cancer.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.