Integrated ensemble CNN and explainable AI for COVID-19 diagnosis from CT scan and X-ray images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Accurate radiological image analysis is crucial for COVID-19 detection, even post-pandemic.
• Current deep learning models lack explainability, hindering clinical adoption.
• Need for interpretable AI in medical diagnostics.

Purpose of the Study:

• To develop an accurate and interpretable deep learning model for COVID-19 detection in radiological images.
• To integrate explainable AI techniques with ensemble Convolutional Neural Network (CNN) models.
• To improve clinician trust and acceptance of AI in diagnosing COVID-19.

Main Methods:

• Ensemble of CNN models (DenseNet169, ResNet50, VGG16) trained on large X-ray (COVIDx CXR-3) and CT datasets.
• Integration of explainable AI techniques: LIME, SHAP, Grad-CAM, Grad-CAM++.
• Cross-dataset evaluation using additional public datasets for robustness.

Main Results:

• High performance on X-ray dataset: 99.00% sensitivity, 99.00% specificity, 99.00% accuracy, 0.99 F1-score, 0.99 AUC.
• Strong performance on CT dataset: 96.18% sensitivity, 96.18% specificity, 96.18% accuracy, 0.9618 F1-score, 0.96 AUC.
• Explainable AI provided transparent insights into model decision-making.

Conclusions:

• The proposed ensemble model with explainable AI significantly improves accuracy and interpretability for COVID-19 detection.
• This approach bridges the gap between AI performance and clinical usability.
• Enhanced disease diagnosis and increased clinician acceptance are anticipated.