Enhancing Automated Prostate Cancer Diagnosis in Histopathology Whole-Slide Images using Weighted-Average Ensembling

Prostate cancer (PCa) is considered the second most common type of cancer among males worldwide, with a very high mortality rate. The International Society of Urologic Pathologists (ISUP) grading system is one of the most reliable indicators for assessing the severity of PCa. ISUP can be used on histopathology tissue samples to evaluate the growth patterns of cancerous cells using a scale ranging from 1 to 5, to reflect the aggressiveness of PCa. Many deep learning (DL) techniques have been applied to enhance PCa diagnosis in histopathology tissue samples. The results have been promising and surpassed traditional human methods. Despite their relatively good performance compared to traditional methods, much effort still needs to be exerted in this area to improve the reliability of DL models and develop more accurate and effective models. Most existing techniques heavily rely on large convolutional neural network (CNN) models, which often have issues regarding long learning times, large dataset requirements, and high computational and storage costs. In this study, we propose a PCa diagnosis system that relies on lightweight CNN models, specifically EfficientNet-B0 and MobileNetV3-Large models, combined with a weighted-average ensembling technique for predicting ISUP grades on Prostate cANcer graDe Assessment (PANDA) dataset. The final ensemble model achieved a Quadratic Weighted Kappa (QWK) score of 0.89 and an accuracy of 70%, outperforming other DL approaches for automated ISUP grading on the PANDA dataset.