Seminars at CBA

Diagnosing prostate cancer is currently based on histopathological assessment of biopsies under a microscope. Pathology departments worldwide are struggling with an increasing number of biopsies and a shortage of pathologists, a problem that can potentially be alleviated by the use of computational pathology and deep neural networks (DNN). With the aim of developing a system for assessing prostate biopsies, we digitized approximately 9000 biopsies from over 1400 men, mainly collected in Stockholm as part of the 2012-2014 STHLM3 diagnostic study. We designed pre-processing algorithms to extract training data according to hand-drawn annotations on the slides, and trained an ensemble of 60 InceptionV3 DNNs for classifying tissue tiles as either benign or malignant tissue of different grades. An ensemble of gradient boosted trees was then trained on the outputs of the DNNs to predict cancer presence, grade and length for each entire biopsy. The model training was performed on large-scale compute clusters with Tesla V100 GPUs. In terms of detecting biopsies with cancer, the system achieved AUC values of 0.997 and 0.986 on internal and external test sets, respectively. The correlation between cancer length relative to that assigned by the study pathologist was 0.96 and 0.87 on the two test sets. In terms of assigning cancer grades, the system achieved a mean pairwise Cohen’s kappa of 0.62 with 23 expert uropathologists on an additional test set (Imagebase). The performance was comparable to the experts (0.60-0.73). Clinical application of a DNN system could reduce pathology workload by reducing the assessment of benign biopsies and by automating the task of measuring cancer length in positive biopsy cores. An AI system with expert-level grading performance might contribute a second opinion, aid in standardising grading, and provide pathology expertise in parts of the world where it does not exist. Article: https://www.thelancet.com/journals/lanonc/article/PIIS1470-2045(19)30738-7/fulltext Demo: https://tissuumaps.research.it.uu.se/sthlm3/index.html