Abstract / Summary
Abstract To the best of our knowledge, existing benchmarks for bias mitigation in skin cancer classification focus exclusively on image-based, leaving a gap in the tabular-multimodal setting, where image-derived probabilities are combined with structured clinical data. To address this gap, this paper proposes an experimental framework to benchmark and integrate multiple bias mitigation algorithms into automated skin cancer classification pipelines based on a Mixed Fusion architecture. Using a multimodal fusion architecture originally developed for interpretability, we extend the framework to assess diagnostic bias across image data and metadata. Experiments were conducted on three datasets: PAD-UFES-20, HIBA, and MIDAS, incorporating sensitive demographic attributes to enable a fairness evaluation. Bias mitigation techniques were systematically applied at the pre-processing, in-processing, and post-processing stages, as well as in combination, using Multi-layer Perceptron (MLP), K-Nearest Neighbors (KNN), and Decision Tree (DT) classifiers. Performance and fairness were jointly analyzed using multiple metrics. According to the A-TOPSIS ranking, the MLP achieved the best results on DB-PAD-UFES-20, with a Balanced Accuracy (BACC) of 0.8313 and Statistical Parity (SP) of 0.0050 for gender and 0.0580 for Fitzpatrick skin type. On DB-HIBA, KNN ranked first, reaching a BACC of 0.7922 and SP of 0.0047 for gender. For DB-MIDAS, MLP again performed best, achieving a BACC of 0.6911 and SP of 0.0096 for gender. Post-processing methods consistently demonstrated effectiveness across datasets. By providing a methodological foundation with baseline results, this work paves the way for future research to develop more reliable clinical diagnosis that incorporate other bias mitigation techniques.