Identification of metastatic primary cutaneous squamous cell carcinoma utilizing artificial intelligence analysis of whole slide images

Cutaneous squamous cell carcinoma (cSCC) harbors metastatic potential and causes mortality. However, clinical assessment of metastasis risk is challenging. We approached this challenge by harnessing artificial intelligence (AI) algorithm to identify metastatic primary cSCCs. Residual neural network-architectures were trained with cross-validation to identify metastatic tumors on clinician annotated, hematoxylin and eosin-stained whole slide images representing primary non-metastatic and metastatic cSCCs (n = 104). Metastatic primary tumors were divided into two subgroups, which metastasize rapidly (≤ 180 days) (n = 22) or slowly (> 180 days) (n = 23) after primary tumor detection. Final model was able to predict whether primary tumor was non-metastatic or rapidly metastatic with slide-level area under the receiver operating characteristic curve (AUROC) of 0.747. Furthermore, risk factor (RF) model including prediction by AI, Clark's level and tumor diameter provided higher AUROC (0.917) than other RF models and predicted high 5-year disease specific survival (DSS) for patients with cSCC with 0 or 1 RFs (100% and 95.7%) and poor DSS for patients with cSCCs with 2 or 3 RFs (41.7% and 40.0%). These results indicate, that AI recognizes unknown morphological features associated with metastasis and may provide added value to clinical assessment of metastasis risk and prognosis of primary cSCC.