Cancer Detection in Radical Prostatectomy Histology using Convolutional Neural Networks
Abstract
Radical prostatectomy (RP) is a common treatment for prostate cancer. We used RP whole-mount tissue samples from 68 patients stained with haematoxylin and eosin to create cancer maps using four pretrained networks: AlexNet, NASNet, VGG16 and Xception, to classify regions of interest (ROIs) as cancer or non-cancer. Models were trained on either raw images or as tissue component maps (TCMs) containing nuclei, lumina and stroma/other components generated from a trained U-Net.
All models performed similarly; however, VGG16 trained on raw images performed with the highest area under the receiver operating characteristic curve (AUC) of 0.994 (95% confidence interval 0.992-0.996). Ensemble models using models trained on raw images performed with the lowest false positive rates. All models had high false negative rate for Gleason 5 cancer and those trained on raw images performed with higher AUC than models trained on TCMs.