Histologic Distinction Between Clear Cell Sarcoma and Melanoma Using Supervised and Deep Learning

Clear cell sarcoma overlaps histologically and immunophenotypically with melanoma. Molecular testing for EWSR1 rearrangement aids in their distinction but may not be readily available. We aim to construct machine learning-based classifiers using supervised and deep learning. We digitized hematoxylin-and-eosin-stained slides from clear cell sarcomas and melanomas with confirmatory testing, constructed nuclear morphometric-based and deep learning-based classifiers (using CLAM/ResNet-50, CTransPath, and UNI models), and evaluated their performance using an independent external validation set. Morphometric analysis of 1,954,194 nuclei (1,308,124 from clear cell sarcomas, 646,070 from melanomas; 8700-70,726 [median 32,472] nuclei per slide) yielded two optimal classifiers using single-node decision tree models; both pertained to nuclear perimeter and included interquartile/interdecile range normalized to median. Factors associated with inaccurate predictions included < 10,000 nuclei/sample and altered morphology in clear cell sarcoma due to therapy. In the external validation set, comparable to the prediction accuracies by four pathologists (median 80%; range 60%-100%), the two nuclear morphometric-based classifiers achieved accuracies of 80%-90%, and the optimal deep learning-based classifier CLAM/CTransPath showed an accuracy of 90%. In conclusion, we have derived interpretable nuclear morphometric- and deep learning-based classifiers to distinguish clear cell sarcoma from melanoma. Quantitative morphometric analysis with machine learning holds the potential as a diagnostic adjunct.