Accurate segmentation of prostate cancer histomorphometric features using a weakly supervised convolutional neural network

Abstract

Prostate cancer (PCa) arises from the glandular epithelium. To confirm the presence of cancer, a pathologist uses stained tissue samples taken either from biopsy or prostatectomy. Due to the relationship between epithelium and cancer, special consideration is given to regions identified as epithelium. Histomophometric techniques have long been used to identify areas of epithelium within the tissue for automated detection and classification pipelines; however, they are often rigid in their implementation and their performance suffers on large datasets where variation in staining, imaging, and preparation is difficult to control. The recent development and popularity of deep learning methods for image processing and segmentation offer great promise for developing robust classification pipelines for such ends; however, they require large labeled datasets for training.The goal of this study was to combine weakly labeled datasets generated using histomophometric techniques, and high‐quality labeled datasets from human observers to train a convolutional neural network. In doing so we developed a pixel‐wise segmentation algorithm for classification of stromal, epithelium, and lumen (SEL) regions for use on both biopsy core and whole‐mount bright‐field H&E stained tissue. We provide evidence that by simply training a deep learning algorithm on weakly labeled data, we can improve the robustness of the classification. Finally, we show that not only does this method carry primary improvement on SEL labeling within tissue, but that the information provided by the deep learning generated labels improves cancer classification in a higher‐order algorithm over the histomorphometric labels that it was trained on.Support or Funding InformationFunding was provided by the State of Wisconsin Tax Check‐off Program for Prostate Cancer Research (RO1CA218144 and R01CA113580) and the National Center for Advancing Translational Sciences (NIH UL1TR001436 and TL1TR001437).This abstract is from the Experimental Biology 2019 Meeting. There is no full text article associated with this abstract published in The FASEB Journal.