Fig. 4

Comparison of indolent and aggressive prediction in core needle biopsy whole slide images (WSIs). Comparison between two trained deep learning models with and without fully supervised learning ([TL-Colon poorly ADC (x20, 512) and WS] and [TL-Colon poorly ADC (x20, 512) and FS+WS]). In (A), Gleason pattern 3 adenocarcinoma (D) was observed in all fragments. The heatmap images show indolent prediction outputs (B, C, E, F). As compared to the weakly supervised (WS) model (B, E), fully supervised (FS) and WS model predicted indolent morphology (Gleason pattern 3) more precisely (F) and indolent predicted area was almost same as pathologist’s marking with blue ink-dots. In (G), the pathologist had missed identifying Gleason pattern 3 adenocarcinoma in (J). WS model did not predict the presence of adenocarcinoma in the same area (K). FS+WS model predicted precisely indolent (Gleason pattern 3) area (L). In (M), infiltrating single cell adenocarcinoma (Gleason pattern 5) (P) was predicted correctly as aggressive (Q) by WS model. FS+WS model predicted infiltrating adenocarcinoma as aggressive more precisely (R). The heatmap uses the jet color map where blue indicates low probability and red indicates high probability