A newly developed deep learning tool accurately predicts metastatic risk in cutaneous squamous cell carcinoma (cSCC), outperforming current clinicopathological and molecular methods for skin cancer.
cSCC is the second most prevalent skin cancer. It carries a 2–5% risk of metastasis, often resulting in poor outcomes, and some of the commonly used risk stratification models have limited predictive accuracy.
To overcome these limitations, researchers from Queen Mary University of London and collaborators from Glasgow’s Queen Elizabeth Hospital, Cheltenham General Hospital and Southampton General Hospital developed cSCCNet – a deep learning model to assess metastatic risk in cSCC.
Automated precision through deep learning
The two-step model was trained on digital histopathology slides from 227 cases of cSCC.
The first model for automated area selection identified the most relevant tumour regions within a whole-slide image, including tumour tissue, inflammatory cells and surrounding stroma, while excluding irrelevant or artefact areas, ensuring that only diagnostically significant regions are analysed.
The second model for the prediction of metastatic risk analysed the selected image tiles, assigned probability scores to each, and aggregated them to generate a tumour-level classification of high or low risk.
Heatmaps generated by the model highlighted the areas that contributed most to its predictions, thereby improving interpretability and supporting validation against known histopathological features, such as poor differentiation, acantholysis and desmoplasia.
In the testing cohort (n=40), cSCCNet achieved an area under the curve of 0.95 and 95% accuracy (CI 0.87–1), outperforming both a 20-gene expression profile and standard clinical classifications.
The model automatically identified tumour regions before assessing metastatic risk and was found to be an independent predictor of metastasis in multivariate analysis, identifying features beyond known clinicopathological markers.
Potential in cSCC clinical practice
The researchers emphasised that validation in prospective, multi-centre and international cohorts is required, and that multidisciplinary collaboration will be critical in developing the model.
The next steps will focus on enhancing explainability, applying cSCCNet to diverse cSCC subtypes and integrating molecular and clinical data into a multimodal tool, the researchers said.
This study highlights the potential of computational histopathology in assessing cSCC. Ultimately, such models could inform treatment decisions, lead to more efficient uses of healthcare resources and optimise care for patients with this prevalent cancer.
Reference
Peleva E et al. Enhanced metastasis risk prediction in cutaneous squamous cell carcinoma using deep learning and computational histopathology. Npj Precision Oncology 2025;9:308.