EGAE

EGAE employs ensemble-based genetic algorithms to generate segmentation-based explanations that identify informative regions in medical images for melanoma cancer detection.

Key Features:

• Heuristic determination of chromosome sparsity: Heuristically determines chromosome sparsity to configure genetic algorithm representations for image segmentation.
• Consecutive execution of multiple genetic algorithms: Executes multiple genetic algorithms consecutively with different numbers of superpixels, producing variations in chromosome length to explore segmentation alternatives.
• Ensemble via consensus and majority voting: Ensembles outputs from multiple GAs using consensus and majority voting to integrate segmentation results.
• Euclidean distance accuracy metric: Quantifies explanation accuracy by computing the Euclidean distance between expert-delineated ground-truth explanations and explainer-generated explanations.
• Benchmarking against LIME: Evaluates explanation accuracy relative to Local Interpretable Model-agnostic Explanations (LIME) on melanoma datasets.
• Automatic identification of informative regions: Uses the ensemble GA process to automatically identify and present informative sections of medical images.

Scientific Applications:

• Melanoma cancer detection: Identifying informative lesions and supporting explanation of AI model decisions in melanoma image datasets.
• Explainable medical imaging: Providing segmentation-based explanations to assess and improve interpretability of image-analysis models.
• Comparative evaluation of explainers: Benchmarking and comparing explanation methods such as EGAE and LIME on annotated medical image datasets.

Methodology:

Three-phase process: heuristic determination of chromosome sparsity; consecutive execution of multiple genetic algorithms with varying superpixel counts yielding different chromosome lengths; ensembling of GA outputs via consensus and majority voting; accuracy measured by Euclidean distance to expert-delineated explanations.