Explainable Artificial Intelligence for Bias Detection in COVID CT-Scan Classifiers.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer-Aided Diagnosis

Background:

• High-performing COVID CT-scan classifiers may inadvertently learn from spurious artifacts, leading to unreliable predictions.
• Explainable Artificial Intelligence (XAI) techniques are crucial for identifying and mitigating such biases in medical AI models.

Purpose of the Study:

• To apply and evaluate various XAI methods for analyzing COVID CT-scan classifiers.
• To assess the robustness of different deep neural network architectures against dataset biases.

Main Methods:

• Utilized multiple XAI techniques including GradCAM, LIME, RISE, Squaregrid, and direct Gradient approaches (Vanilla, Smooth, Integrated).
• Evaluated deep neural network architectures such as VGG16 and DenseNet for COVID CT-scan classification.

Main Results:

• VGG16 demonstrated higher susceptibility to biases from spurious artifacts compared to the more robust DenseNet.
• Small variations in validation accuracy significantly altered explanation heatmaps for DenseNet, indicating sensitivity to learned biases.
• Despite high overall performance metrics (Accuracy, F1, AUC 80-90%), XAI analysis revealed underlying biases in the classifiers.

Conclusions:

• XAI methods are essential for uncovering hidden biases in high-performing medical AI models, even when traditional metrics suggest reliability.
• DenseNet appears more robust against spurious artifacts than VGG16 in COVID CT-scan classification tasks.
• Careful interpretation of XAI-generated heatmaps is necessary to ensure trustworthy AI deployment in clinical settings.