Fig. 7: Slide-level testing results for external STS slide cohort.

A Slide-level ROC curves and AUC scores of externally-sourced CCS, aRMS, and eRMS slide set. B Likelihood distributions resulting from CNN classification of tiles from pathologist-diagnosed CCS slides, with the top (All) representing softmax likelihood scores of all testing images, middle (CNN Agreement) representing likelihood distributions when pathologist diagnosis and softmax-assigned CNN diagnosis match, and bottom (CNN Misdiagnosis) represents likelihood distributions when pathologist diagnosis and softmax-assigned CNN diagnosis are mismatched. Minimum and median correct softmax tile scores were 0.4328 and 0.6575, respectively. C As per (B), but for pathologist-diagnosed aRMS slides. Minimum and median correct softmax tile scores were 0.4083 and 0.5458, respectively. D As per (B), but for pathologist-diagnosed eRMS slides. Minimum and median correct softmax tile scores were 0.4588 and 0.6685, respectively.