Bias Mitigation in Skin Disease Classification

The use of artificial intelligence (AI) in diagnosing skin diseases presents a significant opportunity to enhance healthcare accessibility. However, the effectiveness of AI-based diagnostic systems is often compromised by several challenges, particularly those related to fairness and representation. One prominent issue is the limited diversity in real-world datasets, which can lead to substantial classification biases. This study addresses these challenges by analyzing a dataset collected from an Italian hospital. The dataset exhibits limited data availability, resulting in inadequate representation—especially for darker skin tones. Furthermore, the dataset primarily consists of non-dermoscopic, consumer-grade images, which often suffer from quality issues such as inconsistent lighting and blurriness. These factors collectively complicate the development of accurate and fair AI models for skin disease diagnosis. To address these issues, this research proposes a novel diagnostic pipeline designed to improve both accuracy and fairness in real-world scenarios. The proposed pipeline consists of two main stages: (1) data pre-processing and augmentation, wherein images that better represent darker skin tones are generated using a state-of-the-art diffusion model, and (2) disease classification through deep learning techniques. The efficacy of the proposed methodology is demonstrated through comprehensive validation on real-world data, highlighting significant improvements in both reliability and fairness across various skin disease classifications.