Prediction of pathological grade in prostate cancer: an ensemble deep learning-based whole slide image classification model

Abstract Prostate cancer is a highly prevalent tumor among men and probably histologically represents one of the most common malignant tumors in the world. The automated analysis of tissues in histopathology has been advancing with higher accuracy rates using modern slide imaging techniques in combination with deep learning approaches. Today, pathologists can examine tissue samples in detail on high-resolution digital images thanks to whole slide imaging (WSI) images. This helps to accurately diagnose and evaluate the grades of prostate cancer. Tasks for analyzing WSI images with deep learning methods rely heavily on WSI images where tumor regions are marked by pathologists. This poses the challenge of marking WSI images for pathologists. In this study, we developed deep learning-based models with transfer learning support to classify prostate cancer grade in WSIs based on slide-level labels without the need for pathologists to mark the images. For this aim, we collected a real-life dataset which contains WSI image data of prostate cancer patients obtained from Kocaeli University, Department of Medical Pathology. The collected WSI data were labeled with prostate cancer grades only which are slide-level annotations, without the identification of regions of interest (ROIs) such as tumor regions within the tissue section. To demonstrate the contribution of the proposed, WSI-based classification models, they are compared with patch-based models regarding accuracy, precision, recall, and F1 score. Experimental results show that the proposed model provides high classification accuracy for prostate cancer detection over WSI images.