Fig. 3.

Summary of approach for tile-based deep learning models. a Image tiles are extracted from whole-slide images after any slide processing and buffered into TFRecords. During training, tiles are read from TFRecords, augmented, stain normalized, standardized, and batched. In this weakly-supervised approach, models are trained using ground-truth labels for each tile determined from the label of the corresponding slide. b When evaluating a tile-based model, image tiles do not need to be buffered into TFRecords. Image tiles can be extracted from slides, processed and stain normalized, and predictions are generated for each tile from a slide. The final slide-level prediction is the average prediction from all tiles. Tile-level predictions can be visualized as a heatmap