A Dual-Modality Deep Learning Framework for COVID-19 Detection with Interpretability Using Chest X-Ray and CT Imaging

The present study presents a comprehensive deep learning framework for automatically diagnosing COVID-19 from chest X-rays and CT scans. A balanced, multi-source dataset comprising three groups, which include normal cases, COVID-19 cases, and pneumonia cases, is produced by deploying the proposed method. It compares an optimized AlexNet with a customized Convolutional Neural Network (CNN) using transfer learning. The methodology involves combining patient-level data, adding more data, and conducting a detailed analysis that includes statistical significance testing and 5-fold cross-validation. An independent dataset (BIMCV-COVID19+) is used for external validation to assess generalizability. Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps address concerns about ease of understanding by showing the regions used for diagnosis. The study also tests the effectiveness of the EfficientNet and Vision Transformer (ViT) models. With an accuracy of 99.1%, the ViT outperforms earlier models in clinical screening scenarios with constrained processing resources. The current study highlights the benefits of using this AI-assisted method in actual radiology operations and emphasizes the significance of maintaining thorough records of the datasets.