Toward deploying a deep learning model for diagnosis of rhabdomyosarcoma

In the article, “Machine learning for rhabdomyosarcoma histopathology,” Frankel et al1 showed that a convolutional neural network (CNN)-based differential diagnoses model can be developed as a prepathologist screening system to quantify diagnosis likelihood of 3 soft tissue sarcomas, embryonal rhabdomyosarcoma (eRMS), alveolar rhabdomyosarcoma (aRMS), and clear cell sarcoma (CCS), from hematoxylin and eosin–stained histopathology whole-slide images (WSIs).1 The model was trained on 82 aRMS WSIs, 71 eRMS WSIs, and 10 CCS WSIs; validated on 19 aRMS WSIs, 17 eRMS WSIs, and 2 CCS WSI;, and tested on 18 aRMS WSIs, 15 eRMS WSIs, and 3 CCS WSIs, all from an internal cohort. The model achieved receiver operating characteristic area under the curve (AUC) of 0.90 for aRMS, 0.89 for eRMS, and 1.00 for CCS at slide level. The model was also tested on an external cohort composed of 14 aRMS WSIs, 17 eRMS WSIs, and 30 CCS WSIs and achieved an AUC of 0.89 for aRMS, 0.61 for eRMS, and 0.64 for CCS in slide level.