Abstract
ABSTRACT
Accurate diagnosis and risk stratification are central for optimal management of patients with melanoma. Current American Joint Committee on Cancer (AJCC) staging systems inadequately stratify patients with clinically meaningful metastatic potential. Manual histopathologic interpretation compounds this limitation, particularly for diagnostically ambiguous lesions at the benign‐malignant interface, where concordance is highly variable. This review examines how computational pathology and convolutional neural networks (CNNs) can improve histopathologic diagnosis and risk stratification of melanoma, evaluate current image‐based deep learning (DL) approaches, and outline a path toward explainable, multimodal prognostic tools. We review published DL models applied to hematoxylin and eosin whole‐slide images (H&E WSIs) for melanoma diagnosis, subtype classification, and survival prediction, and discuss integration with transcriptomic and spatial proteomic data modalities. Computational pathology, informed by deep learning, can reliably identify melanomas at risk of disease recurrence and progression. These inferences can be enhanced by integration of spatial molecular profiling, which can also provide mechanistic explainability to the H&E‐based DL models. However, limitations in dataset diversity, external validation, model interpretability, and generalizability across populations and image acquisition protocols currently prevent clinical adoption. Outcome‐anchored, multimodal computational pathology pipelines integrating H&E WSIs with spatial multi‐omic profiling offer a biologically grounded and scalable framework for personalized risk stratification in stage I‐III CM, with potential to standardize diagnosis, discover novel prognostic features, and inform individualized treatment strategies.
image