P004 Enhanced metastasis risk prediction in cutaneous squamous cell carcinoma using deep learning and computational histopathology: cSCCNet development and evaluation
Emilia Peleva, Yue Chen, Bernhard Finke, Hasan Rizvi, Eugene Healy, Chester Y. Lai, Paul J. Craig, William Rickaby, Christina Schoenherr, Craig Nourse, Charlotte M. Proby, Gareth J. Inman, Irene M. Leigh, Catherine A Harwood, Jun Wang · British Journal of Dermatology · 2025
Abstract Cutaneous squamous cell carcinoma (cSCC) is the most common skin cancer with metastatic potential, and development of metastases carries a poor prognosis. To address the need for reliable risk stratification, a multidisciplinary team including dermatologists, pathologists and data scientists developed cSCCNet, a deep learning model using digital pathology of primary cSCC to predict metastatic risk. We present an update on the cSCCNet collaboration, including a more robust model development pipeline and incorporation of histopathological analysis including multiplex immunohistochemistry, to improve model explainability. A retrospective cohort of 228 cSCCs is used for model development. cSCCNet automatically selects the tumour area in standard histopathological slides and then stratifies primary cSCC into high- vs. low-risk categories, with heatmaps indicating the most predictive tiles contributing to explainability. On a 20% holdout testing cohort (n = 41), cSCCNet achieved area under the receiver operating characteristic curve 0.91 and 93% accuracy in predicting risk of metastasis, outperforming a gene expression-based tool and histopathological staging (Table). Multivariate analysis including common staging criteria confirmed cSCCNet as an independent predictor for metastasis. Histopathological analysis including multiplex immunohistochemistry suggests that tumour differentiation, acantholysis, desmoplasia and CD3 T-cell infiltration were important in predicting risk of developing metastasis. Although further validation including prospective evaluation is required, cSCCNet has potential as a reliable and accurate tool for metastatic risk prediction.TablePredictive performance of cSCCNet, 20 gene expression profile test and clinicopathological classifications on the holdout testing cohort (n = 41) AUROCAccuracySpecificitySensitivityNPVPPVcSCCNet0.91 (0.82–1)93%96%87%93%93%20-GEP0.77 (0.63–0.92)79%83%71%83%71%UICC8/AJCC80.73 (0.57–0.88)76%84%61%81%67%BWH0.73 (0.57–0.88)76%84%61%81%67%BAD high/very highNA63%44%100%100%48%BAD very high only0.71 (0.55–0.87)73%80%62%80%62%Performance in predicting risk of cutaneous squamous cell carcinoma (cSCC) metastasis, based on cSCCNet prediction, the 20-gene expression profile model (20-GEP), the 8th edition of the Union for International Cancer Control/American Joint Committee on Cancer (UICC8/AJCC8), Brigham and Women’s Hospital classification (BWH) and the British Association of Dermatologists (BAD) cSCC guidelines (based on ‘high/very high’ grades or ‘very high’ grade only). AUROC, area under the receiver operating characteristic curve (with 95% confidence interval); NPV, negative predictive value; PPV, positive predictive value.