Self-Learning for Weakly Supervised Gleason Grading of Local Patterns

Prostate cancer is one of the main diseases affecting men worldwide. The gold standard for diagnosis and prognosis Gleason grading system. In this process, pathologists manually analyze prostate histology slides under microscope, in a high time-consuming subjective task. last years, computer-aided-diagnosis (CAD) systems have emerged as promising tool that could support daily clinical practice. Nevertheless, these are usually trained using tedious prone-to-error pixel-level annotations grades tissue. To alleviate need manual pixel-wise labeling, just handful works been presented literature. Furthermore, despite results achieved on global scoring location cancerous patterns tissue only qualitatively addressed. These heatmaps tumor regions, however, crucial to reliability CAD they provide explainability system's output give confidence model focusing medical relevant features. Motivated by this, we propose novel weakly-supervised deep-learning model, based self-learning CNNs, leverages score gigapixel whole slide images during training accurately perform both, patch-level biopsy-level scoring. evaluate performance proposed method, extensive experiments three different external datasets grading, two test sets Grade Group prediction. We empirically demonstrate our approach outperforms its supervised counterpart large margin, well state-of-the-art methods Particularly, brings an average improvement Cohen's quadratic kappa (?) nearly 18% compared full-supervision This suggests absence annotator's bias capability weakly labeled leads higher performing more robust models. raw features obtained from classifier showed generalize better than previous approaches literature