MetaCAM as an ensemble-based class activation mapping framework improves model explainability.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning Explainability

Background:

• Deep learning model predictions require clear explanations in high-criticality fields like medicine.
• Class Activation Maps (CAMs) are popular for explaining Convolutional Neural Networks (CNNs).
• Individual CAM performance varies significantly with experimental parameters.

Purpose of the Study:

• To develop an ensemble-based method, MetaCAM, for combining multiple CAMs.
• To improve the reliability and accuracy of visual explanations from deep learning models.
• To refine salient image regions identified by CAMs for better model interpretation.

Main Methods:

• Proposed MetaCAM, an ensemble method aggregating CAMs based on consensus of top activated pixels.
• Quantitatively determined optimal CAM combinations for MetaCAM experiments.
• Introduced Cumulative Residual Effect (CRE) for summarizing large-scale ensemble experiments.
• Implemented adaptive thresholding to enhance individual CAM performance.
• Evaluated performance using the Remove and Debias (ROAD) pixel perturbation method.

Main Results:

• MetaCAM demonstrated superior performance compared to individual CAMs.
• MetaCAM refined salient regions in images used for model predictions.
• MetaCAM improved ROAD performance to 0.393, outperforming individual CAMs (-0.101 to 0.172).
• Adaptive thresholding was shown to enhance individual CAM performance.

Conclusions:

• Ensembling multiple CAMs with adaptive thresholding significantly improves explanation reliability.
• MetaCAM offers a robust approach for generating trustworthy visual explanations in critical applications.
• The proposed methods provide a quantitative framework for evaluating and enhancing CAM-based explanations.