Evolutionarily Optimized Multi-Scale Gabor Modeling of Directional Lesion Texture in Dermoscopic Images for Interpretable Melanoma Classification

Background: Melanoma is one of the most aggressive forms of skin cancer, making early and accurate diagnosis essential for improving patient outcomes. Methods: In this work, we propose an Evolutionary Gabor-based Melanoma Descriptor (Evo-GMD), a lightweight and interpretable approach designed under the principles of Frugal AI. The method integrates multi-scale Gabor filtering with Differential Evolution to automatically learn discriminative texture patterns using a reduced set of parameters. The proposed approach was evaluated on the PH2 dataset, achieving competitive performance (accuracy above 95%) while maintaining low computational complexity and full interpretability. To further assess its robustness, complementary experiments were conducted on the ISIC 2017 dataset, which presents higher variability, class imbalance, and heterogeneous lesion characteristics. Results: The results reveal that multiple methods—including handcrafted descriptors, convolutional neural networks, and transfer learning models—exhibit significant performance degradation or converge to trivial solutions under these conditions. This behavior highlights that increasing model complexity does not necessarily improve classification performance when data constraints are present. Conclusions: Overall, the findings demonstrate that the proposed method provides a robust and efficient alternative for melanoma classification in low-resource scenarios, where data availability, computational capacity, and interpretability are critical factors.