SmartProg-MEL-integrating dermatopathology and explainable artificial intelligence (AI) to improve prognostic accuracy and risk stratification in cutaneous melanoma

In the era of precision oncology, artificial intelligence (AI) is increasingly recognized for its potential to transform histopathological workflows and prognostic stratification. The study by Bossard et al.¹ introduces SmartProg-MEL, a deep learning–based model developed to predict 5-year overall survival (OS) in patients with primary cutaneous melanoma (stages I–III), using only routine haematoxylin–eosin (HE) or HE-saffron–stained whole slide images (WSIs). By offering an explainable, image-based prognostic tool, SmartProg-MEL aims to complement or surpass traditional clinicopathological risk assessments.

A major strength of this study lies in its robust validation across multiple independent cohorts. The model was trained on a discovery dataset of 342 patients (IHP-MEL-1) and externally validated on two independent sets: IHP-MEL-2 (n = 161) and TCGA (n = 63). It consistently achieved concordance indices of 0.72, 0.71 and 0.69, respectively, with sensitivity ranging from 71% to 100%. These metrics outperform earlier AI prognostic tools,^2-4 which were often limited by small datasets and lack of external validation. Notably, SmartProg-MEL maintained high performance across varying staining protocols and scanner types, supported by robust stain normalization and augmentation techniques—critical for clinical translation.

Clinically, the model's utility is underscored by its ability to refine prognostic classification beyond the current AJCC TNM staging system.⁵ For instance, it identified high-risk patients within stage I melanomas—traditionally considered low risk—and reclassified a significant portion of stage IIB/IIC cases as low risk. Such refined stratification could support tailored decisions around surveillance intensity and adjuvant immunotherapy, potentially sparing low-risk individuals from overtreatment while ensuring timely intervention for those at greater risk.

Equally notable is the emphasis on interpretability. Using attention-based heatmaps and UMAP-based feature clustering, the model highlights morphologic correlates of risk. Features such as nuclear pleomorphism, cellular atypia, architectural disorganization and peritumoral immune infiltrates align with established histopathological prognostic factors. This transparency enhances clinician trust and positions SmartProg-MEL not merely as a ‘black box’ tool, but as an interpretable digital biomarker.

However, several limitations warrant discussion. The retrospective design introduces inherent selection biases, and treatment-related variables during follow-up were not incorporated into the survival models. This is particularly relevant given the growing impact of systemic therapies on melanoma outcomes. Furthermore, although SmartProg-MEL performed well across cohorts, a slight dip in performance in the TCGA set may reflect technical and biological heterogeneity. Another challenge is translating tile-level risk predictions into practical tools that pathologists can readily interpret in clinical workflows.

Despite these caveats, SmartProg-MEL's impact is potentially transformative. By deriving prognostic insights from routine diagnostic slides, it offers a non-invasive, cost-effective addition to clinical decision-making. Future directions should prioritize prospective, multi-institutional validation studies, integration with molecular and clinical variables and exploration of treatment response prediction—particularly for immunotherapy.

In conclusion, SmartProg-MEL represents a compelling advance in digital pathology. Its strong performance, generalizability and focus on interpretability make it a valuable tool for enhancing melanoma prognosis and guiding individualized patient management. With continued refinement and clinical integration, SmartProg-MEL could play a pivotal role in the future of precision oncology for melanoma.

None.

The authors declare no conflict of interest.

Not applicable.

Not applicable.