BACKGROUND:
Artificial intelligence (AI) is increasingly adopted in cancer screening and clinical decision-making, yet persistent racial bias in dermatological AI systems risks reinforcing global health inequities. ¹-³ Melanoma survival disparities exemplify this challenge: Black patients experience an estimated five-year survival rate of approximately 66%, compared to 90% among White patients, ⁴ largely due to late-stage diagnosis. A structural driver of this inequity is dataset imbalance, as widely used repositories such as HAM10000 are predominantly sourced from Western populations and demonstrate reduced diagnostic accuracy for Fitzpatrick skin types IV–VI. ⁵ These limitations conflict with global policy principles articulated by the World Health Organization (WHO) and UNESCO, which emphasize fairness, inclusivity, and data sovereignty in AI for health.

METHODS:
This study proposes a federated learning (FL) framework using the Federated Averaging (FedAvg) algorithm⁶ to train a ResNet-50 classifier without centralizing patient data. ⁷ To reflect the fragmented and resource-constrained data ecosystems common in low- and middle-income countries (LMICs), non-independent and identically distributed (non-IID) partitions ⁸ were simulated through re-sampling of public datasets (HAM10000/ISIC) ⁹. Model optimization prioritized sensitivity and false negative rate (FNR) for darker skin phenotypes, aligning technical performance with equity-relevant clinical outcomes.¹⁰

RESULTS:
As a proof-of-concept, the federated model achieved sensitivity comparable to centralized training while preserving institutional data ownership and privacy. The framework demonstrates how collaborative AI development can occur across distributed health systems without extractive data practices.

CONCLUSIONS:
Federated learning represents a policy-aligned architecture for equitable AI deployment in oncology, particularly in LMIC contexts. By supporting privacy protection, cross-border collaboration, and fairness auditing, this approach aligns with WHO and UNESCO guidance and offers a pathway for regulators and health systems to reduce preventable diagnostic failures contributing to melanoma mortality in underrepresented populations.