Abstract Squamous cell carcinoma (SCC) is a common malignancy that arises in diverse organs and often exhibits overlapping histological and immunophenotypic features, making accurate determination of the tissue of origin challenging with conventional pathology. To address this, we developed Deepath-SCC, a deep learning model for pan-SCC tissue origin identification directly from hematoxylin and eosin-stained whole-slide images. A retrospective cohort of 4217 whole slide images across nasopharyngeal, head and neck/esophageal, lung, cervical, and urothelial carcinomas was assembled for model training and validation. In the internal test set, Deepath-SCC reached accuracy of 91.2%, with micro area under the receiver operating characteristic curves (AUROC) reaching 0.986 (95% CI: 0.984–0.989). Among high-confidence predictions (similarity score ≥0.7914), the overall accuracy increased to 96.2%, including 96.4% for primary SCCs and 94.4% for metastatic SCCs. In an external test set, Deepath-SCC achieved an accuracy of 86.1% with an AUROC of 0.972 (95% CI: 0.961–0.982). These results provide initial evidence supporting the feasibility of deep learning–based digital pathology for tissue-of-origin prediction in pan-SCC. Deepath-SCC represents an efficient, and cost-effective computational approach that may complement existing diagnostic workflows, particularly in challenging or resource-limited clinical settings.
Lu et al. (Fri,) studied this question.