Enhancing Early Detection of Metastatic Cutaneous Squamous Cell Carcinoma (CSCC): Integrating AI With Histopathological Assessments

Cutaneous squamous cell carcinoma (CSCC) patients at high risk for metastasis are insufficiently identified with current staging systems. Advances in digital pathology and artificial intelligence (AI) might assist by extracting detailed and reproducible predictive features from haematoxylin and eosin slides. We evaluated a multi-step convolutional neural network (CNN) as an assistive tool to provide detailed complementary histopathological variables towards identifying high-risk CSCC. Using a nested case-control design, we studied patients diagnosed with primary CSCC in the Netherlands from 2007 to 2018, with metastatic patients as cases and non-metastatic patients as controls. The dataset was divided into a development set (130 patients) and an evaluation set (244 patients). Four elaborative variables were derived from a CNN model for object detection and semantic segmentation, complementing six dermatopathologist-scored histopathological variables. Dermatopathologists involved were blinded to the outcomes. We assessed the efficacy of these variables using multivariable logistic regression (MR) models and odds ratios (OR) for metastatic CSCC on the evaluation set. The MR model fitting was assessed using the pairwise concordance index (C-index). The combined dermatopathologist-AI model yielded the highest C-index (0.92 [0.87-0.95]). Significant variables in the combined model included model-derived tumour area (OR 1.35 [1.00-1.84]) which complemented scored tumour diameter (OR 1.54 [0.75-3.17]) and model-derived nuclei density (OR 3.14 [1.08-9.17]) as a counterpart of scored tumour differentiation grades (OR 10.6 [3.01-37.0] and 11.5 [2.95-44.5]). The CNN model can derive detailed and reproducible histopathological variables associated with metastatic risk in CSCC, complementing the current pathologist-based assessment and enhancing the identification of high-risk CSCC.