From underrepresentation to evidence: Pipsqueak dataset and explainable analysis of skin tone bias in melanoma detection

Cutaneous melanoma presents a profound healthcare challenge, particularly for individuals with darker skin tones, as current dermatological image classification systems suffer from severe underrepresentation of diverse skin tones in training datasets. This research uses MultiExCam, our novel architecture, to quantitatively demonstrate the systemic bias in melanoma detection across different skin tones. Our contributions are threefold: first, we comprehensively analyze major dermatological image datasets, analyzing the severe underrepresentation of Fitzpatrick skin types V–VI through fairness metrics; second, we introduce Pipsqueak, a meticulously curated dataset of melanocytic lesions in darker skin tones, comprising 16 core images expanded to 80 through controlled augmentation, which demonstrates the profound scarcity of diverse representation in existing resources; and third, through empirical validation, we quantify performance disparities that emerge when models trained predominantly on light skin images are applied to darker skin tones. Our analysis reveals that the imbalance ratio for the Fitzpatrick17k dataset reaches an imbalance ratio of 7.57, indicating severe representational disparity, while the model’s sensitivity for melanoma detection plummets from 67% on HAM10000 to merely 11% on Pipsqueak. Through explainable AI techniques, we demonstrate that models trained on biased datasets fail to identify clinically relevant features on darker skin, frequently attending to surrounding skin rather than lesion characteristics. This work extends the paper in Ruga et al. (2025) and provides crucial evidence for the urgent need to develop more inclusive diagnostic technologies that can effectively serve all populations, regardless of skin tone, and challenges the field to prioritize collection of diverse dermatological data.