Enhancing Diagnostic Equity: A Deep Learning Framework for Fair Skin Lesion Classification

Rajeev Ranjan Dwivedi, Monika Sharma, Amit Sangroya

Keywords: Fairness, Skin Lesion Diagnosis, Bias Mitigation

Abstract: Algorithmic bias remains a critical challenge in dermatological diagnosis, especially as deep learning models often underperform for underrepresented populations. In this work, we present a novel framework that integrates bias mitigation directly into the training process for skin lesion classification. Motivated by the chronic underrepresentation of darker skin tones (Fitzpatrick types V–VI) in standard dermatology resources, our approach employs a composite loss function that jointly optimizes disease classification and skin-tone prediction. By incorporating cosine dissimilarity regularization, the method encourages the learning of disentangled, robust feature representations, while a Gradient Reversal Layer ensures that these features remain invariant to skin tone. Evaluated on both the Fitzpatrick-17k and ISIC datasets, our framework demonstrates significant improvements in fairness and accuracy, paving the way for more equitable diagnostic tools in medical imaging.

Submission Number: 102