Fig. 1: Example true positive (correctly identified as having PTEN loss) cases from internal and external testing cohorts.

Note external probability maps show performance of the algorithm after fine-tuning. Top: internal cohort testing data set; case with complete PTEN loss in tumor cells, multi-resolution probability = 43.4% and dice = 0.738. Middle top: internal cohort testing data set; case with partial PTEN loss, multi-resolution probability = 56.6% and dice = 0.761. Middle bottom: external cohort testing data set; case with complete PTEN loss in tumor cells, multi-resolution probability = 37.7% and dice = 0.552. Bottom: external cohort testing data set; case with partial PTEN loss in tumor cells, multi-resolution probability = 27.7% and dice = 0.347. Cores with probability > 26.5% were classified as “PTEN loss” in multi-resolution approach. Cores with any pixel region classified as PTEN loss in binary mask were classified as “PTEN loss” in pixel-based approach.