Fig. 4: Biopsy Simulation Experiments.

a Example of a WSI with heterogeneous tile-level imCMS classification. b The shapes of simulated biopsy fragments are randomly sampled from annotations of biopsies in the GRAMPIAN and ARISTOTLE cohorts. c Illustration of a biopsy simulation procedure based on a source imaged resection specimen. Biopsy samples are randomly generated by sampling m = 1 biopsy fragments at random in the annotated tumor region of the resection specimen. This random process is repeated 10,000 times for each WSI of a resection dataset to form a simulated biopsy dataset. For each generated biopsy sample, imCMS is applied. The resulting distributions of class probability over a simulated biopsy dataset are shown on the right. d same as c but with m = 3 biopsy fragments. The spatial heterogeneity of tile-level predictions (shown in a) explains the different distributions obtained for each imCMS class across a simulated biopsy dataset. e Classification performance of trained imCMS models on simulated biopsy datasets as a function of the number of tumor biopsy fragments per sample. Dots with lines indicate the mean and standard deviation of the macro-average AUROC obtained by five models for each simulated dataset: [E4a] (left); [E4b] (right). The red dotted line indicates a 3% difference score with the macro-average AUROC obtained on corresponding fully imaged resection specimens. Scale bars represent 2 mm.