Early detection of metastatic risk in primary cutaneous melanoma using weakly supervised learning.

Early identification of melanomas with high metastatic potential is crucial for treatment planning and survival prediction. We investigated whether weakly supervised WSI-based and multimodal learning approaches can identify metastatic risk in primary cutaneous melanoma beyond current clinicopathological predictors. A total of 426 routinely stained whole-slide images (WSIs), from primary melanomas (249 metastatic and 177 non-metastatic) together with corresponding clinicopathological features, were collected and divided to training and validation, and a hold-out test set. WSIs were divided into patches and encoded using Prov-GigaPath, while clinicopathological features were converted to text embeddings using BioMedBERT. We evaluated a weakly supervised transformer-based approach for slide-level metastatic risk prediction from WSIs, both alone and in combination with clinicopathological features, and compared performance to a model based on clinicopathological features alone. Both WSI-based models achieved strong performance (AUC = 0.887 and 0.883) and higher accuracy (0.847 for both) than the clinicopathological-based model (AUC = 0.849, accuracy = 0.753). This benefit was most pronounced in T2 tumors, where early risk stratification is clinically most relevant. These findings demonstrate that weakly supervised WSI-based models capture prognostically relevant information and highlight their potential for early metastatic risk stratification in primary melanoma.