Evolutionarily Optimized Multi-Scale Gabor Modeling of Directional Lesion Texture in Dermoscopic Images for...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Melanoma, an aggressive skin cancer, necessitates early and accurate diagnosis for improved patient outcomes.
• Current diagnostic methods face challenges with complex datasets and varying lesion characteristics.

Purpose of the Study:

• To develop a lightweight, interpretable, and efficient melanoma classification method.
• To evaluate the proposed method's robustness against complex datasets and Frugal AI principles.

Main Methods:

• Proposed an Evolutionary Gabor-based Melanoma Descriptor (Evo-GMD) integrating multi-scale Gabor filtering and Differential Evolution.
• Utilized Frugal AI principles for automatic learning of discriminative texture patterns with reduced parameters.
• Evaluated performance on the PH2 and ISIC 2017 datasets.

Main Results:

• Evo-GMD achieved competitive performance (accuracy >95%) on the PH2 dataset with low computational complexity and high interpretability.
• Demonstrated superior robustness compared to complex models (CNNs, transfer learning) on the challenging ISIC 2017 dataset.
• Highlighted limitations of complex models in data-constrained scenarios.

Conclusions:

• The Evo-GMD offers a robust and efficient solution for melanoma classification, particularly in low-resource settings.
• Emphasized the importance of interpretability and computational efficiency in AI-driven medical diagnostics.
• Advocated for simpler, data-aware AI models when dealing with limited or variable data.