ExAID

ExAID provides explainable, concept-based analyses of deep learning outputs for computer-aided diagnosis of dermoscopic skin lesions to make AI-driven melanoma classification decisions interpretable.

Key Features:

• Multi-modal concept-based explanations: Produces combined textual descriptions and visual maps that explain model predictions from dermoscopic images.
• Concept Activation Vectors (CAVs): Uses CAVs to map human-understandable concepts onto representations learned by deep neural networks.
• Concept localization maps: Generates visual maps that highlight image regions associated with learned concepts.
• Educational mode (dataset-level explanation statistics): Computes dataset-level explanation statistics for model analysis and educational investigation.
• Concept detectors with reported accuracy: Employs concept detectors that achieved up to 81.46% accuracy, reported as comparable to supervised end-to-end networks.

Scientific Applications:

• Melanoma classification from dermoscopic images: Supports melanoma classification by providing interpretable explanations for deep learning predictions on dermoscopic images.
• Model validation and error analysis: Aids clinicians and researchers in understanding incorrect predictions and refining diagnostic accuracy.
• Evaluation and benchmarking: Enables evaluation of explanatory coherence on publicly available dermoscopic image datasets.

Methodology:

Identifies and explains concepts learned by deep learning models using Concept Activation Vectors (CAVs), concept localization maps, and concept detectors to construct fine-grained textual explanations supplemented with visual information.