Fig. 2

Overview of the image pre-processing, model optimisation and performance evaluation. a. Standardised WSI preprocessing pipeline from retrieval of WSI at 40 × magnification to the cancer-detected tumour tiles from the WSI. b. Schematic overview of the image modelling strategy, including the deep CNN feature extractor and attention module. Model optimisation, hyperparameter tuning, and model selection were performed by cross validation (CV). In each CV training round, the feature extractor and attention module were trained from cancer tiles in the CV training set. In each CV validation round, the features extractor and attention model were re-optimised and subsequently, the CV validation set was evaluated. c. Two cut-offs were further derived from the slide-level prediction scores, which categorised the prediction scores into three-level predicted grades. The cut-offs were optimised by maximising the agreement between the predicted grade and clinical NHG. We further evaluated the prognostic performance of the predicted grade on recurrence-free survival