EAMAPG: Explainable Adversarial Model Analysis via Projected Gradient Descent.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) models demonstrate high performance in medical image analysis but lack transparency.
• Interpretability of DL models is crucial for clinical adoption and trust.

Purpose of the Study:

• To develop and evaluate a novel method for enhancing the interpretability of DL models in medical image analysis.
• To identify key image features that influence DL model decisions using adversarial examples.

Main Methods:

• Utilized projected gradient descent (PGD) to generate adversarial examples by introducing perturbations that cause misclassification.
• Applied the adversarial generation method to analyze medical images from Brain Tumor, Eye Disease, and COVID-19 datasets.
• Evaluated six common convolutional neural network (CNN) models, focusing on top-performing ones like DenseNet121, InceptionV3, and ResNet101.

Main Results:

• Adversarial perturbations significantly increased model loss (p < 0.05), confirming successful adversarial generation and highlighting model vulnerabilities.
• Demonstrated the method's effectiveness across diverse medical imaging datasets (Brain Tumor, Eye Disease, COVID-19).
• Achieved high performance metrics for selected models: DenseNet121 (AUC 1.00), InceptionV3 (AUC 0.99), and ResNet101 (AUC 1.00).

Conclusions:

• The proposed adversarial generation method offers a novel approach to improve DL model interpretability in medical imaging.
• This technique provides a more intuitive understanding of DL model decisions, bridging the gap between AI capabilities and clinical application.
• The findings support the practical use of interpretable AI in clinical settings, aiding radiologists in decision-making.