The potential of convolutional neural network diagnosing prostate cancer

Abstract

Introduction: There were more than 1,276,106 new cases of prostate cancer (PC) in 2018 worldwide (GLOBOCAN). Early and precise diagnosis leads to cure chances up to 90%. Digital rectal examination and PSA serum levels are employed for prostate cancer screening. If both exams are suspicious for cancer, the patient will be submitted to a prostate biopsy. Histological diagnosis and grading are crucial to the proper manage of the patients and are not always easy to evaluate, demanding experience of pathologists. To test the possibility to adopt artificial intelligent to diagnose PC, we studied a set of prostate biopsy sample images that were input to a specifically constructed convolutional neural network. Purpose: Evaluate the potential of the convolutional neural network for the classification of cancer and non-cancer patches extracted from prostate biopsy images. Methods: Thirty-two prostate cancer biopsy images were obtained and reviewed by a single uropathologist and then transformed into 2594 fragments to feed the CNN. The methodology has been divided into clinical approaches—to extract patches—and computational approaches—the CNN implementation. Results: The k-fold three-way cross-validation method was used, resulting in a 98.3% output accuracy in distinguishing cancer from non-cancer. Conclusion: The presented method proved to be robust and trustworthy comparing with an expert pathologist report.