Figure 2.

Performance evaluation on the test set of gland classification dataset and spatial malignancy map

(A and B) Area under receiver operating characteristics curve (AUROC) (A) and average precision (AP) (B) calculated over precision versus recall curve together with 95% confidence intervals (obtained using the percentile bootstrap method²²) are presented for the single-resolution models, three-resolution models, and a four-resolution model.

(C and D) Example manual annotations and spatial malignancy map produced using a deep-learning-based pipeline, respectively. Arrow heads show tissue components detected by the pipeline but not manually annotated.