The rapid spread of COVID-19 necessitates swift and accurate diagnostic methods. While RT-PCR is the gold standard, its limitations in sensitivity and availability have led some countries to utilize chest imaging (CT and X-ray) as a first-line diagnostic tool. However, manual interpretation of CT scans by radiologists is time-consuming and prone to errors, especially when differentiating COVID-19 from similar pneumonias. This study addresses these challenges by developing and evaluating an AI-based system for automated COVID-19 diagnosis from chest CT scans. The AI approach offers the potential for high efficiency, repeatability, and large-scale deployment, alleviating the strain on healthcare systems during outbreaks. Existing AI systems for COVID-19 diagnosis from CT scans have used relatively small datasets or focused on simpler two-class classifications. This research aims to overcome these limitations by utilizing a significantly larger, more clinically representative multi-center dataset and employing a multi-class classification approach. Furthermore, it aims to directly compare the performance of CT and chest X-ray (CXR) using paired data and to provide a detailed interpretation of the AI system's decision-making process to enhance transparency and understanding.

Several studies have explored AI-based systems for COVID-19 diagnosis using CT scans. Zhang et al. developed a system with an AUC of 0.9797, but relied on lesion segmentation, which is time-consuming and not always accurate. Li et al. achieved an AUC of 0.96 but their method had high memory demands. Other studies focused on slice-level analysis or simpler two-class classifications. In contrast, this study utilizes a considerably larger dataset, encompassing multiple pneumonia types and healthy controls, to enable a more robust and clinically relevant multi-class classification.

The researchers developed a deep learning-based AI system comprising five key parts: (1) lung segmentation network; (2) slice diagnosis network; (3) COVID-infectious slice locating network; (4) visualization module for interpreting attentional regions; and (5) image phenotype analysis module. The system was trained on a large multi-center dataset of 11,356 CT scans from 9025 subjects, including COVID-19, community-acquired pneumonia (CAP), influenza, and non-pneumonia cases. The dataset was divided into training and test cohorts, with additional independent test cohorts from publicly available databases (CC-CCII and MosMedData). The lung segmentation network used a U-Net architecture, while the slice diagnosis network used a ResNet152 backbone pre-trained on ImageNet. A task-specific fusion module combined slice-level predictions into volume-level diagnoses for various diagnostic tasks. Guided Grad-CAM was used to visualize the AI system's attentional regions, and radiomics analysis was performed to extract and analyze phenotypic characteristics of these regions. A reader study involving five experienced radiologists compared the AI system's performance to human experts on three tasks: differentiating pneumonia from healthy, COVID-19 from CAP, and COVID-19 from influenza. The performance metrics included AUC, sensitivity, specificity, and accuracy.

The AI system demonstrated high diagnostic accuracy, achieving an AUC of 97.81% for multi-way classification on the test cohort (3199 scans), 92.99% on the CC-CCII dataset, and 93.25% on the MosMedData dataset. In the reader study, the AI system outperformed all radiologists in the more challenging tasks of differentiating COVID-19 from CAP and influenza, achieving AUCs of 0.9727 and 0.9585, respectively. The AI system was significantly faster than the radiologists (2.73 s vs. 6.5 min on average). The analysis of different subject subsets (by gender, age, and disease stage) revealed some performance variations, suggesting potential age and gender-specific factors influencing the AI's ability to detect COVID-19. Comparison of CT and CXR performance using paired data showed that CT significantly outperformed CXR in diagnosing COVID-19, although CXR did show some diagnostic value. Guided Grad-CAM highlighted different attentional regions for different types of pneumonia, suggesting the AI system focused on specific visual features to differentiate between them. Radiomics analysis of the attentional regions revealed imaging characteristics consistent with previous literature on COVID-19 pathogenesis and morphology.

The high diagnostic accuracy of the AI system, surpassing that of experienced radiologists in challenging tasks, highlights its potential as a valuable clinical tool for rapid COVID-19 diagnosis. The speed advantage of the AI system offers significant improvements in diagnostic efficiency, potentially mitigating the bottleneck of manual interpretation during outbreaks. The analysis of different subject subsets provides insights into the factors influencing diagnostic performance and could inform targeted strategies for improving diagnostic accuracy in specific populations. The comparison of CT and CXR performance provides valuable information for resource allocation and clinical practice. The interpretability analysis using Grad-CAM and radiomics helps to understand the AI system's decision-making process, building trust and providing insights into the imaging characteristics of different pneumonia types. Future studies could explore using larger and more diverse datasets, investigate more advanced fusion techniques for combining CT and CXR information, and further refine the interpretability techniques to improve clinical usability.

This study demonstrates the development and validation of a highly accurate and efficient AI system for COVID-19 diagnosis using chest CT scans. The system's performance surpasses that of experienced radiologists in challenging diagnostic tasks, highlighting its potential to significantly improve the efficiency and accuracy of COVID-19 diagnosis. The interpretability analysis enhances transparency and clinical understanding. Future work could focus on incorporating additional clinical data, improving the accuracy of lesion segmentation for radiomics analysis, and exploring the integration of the AI system into clinical workflows to maximize its impact on patient care.

The study's retrospective nature and reliance on data from a specific geographical region may limit the generalizability of the findings. The quality of the CXR data (localizer scans) might have underestimated the true performance of CXR in comparison with CT. While the interpretability analysis provides valuable insights, it does not fully explain the AI system's complex decision-making process. Future work should address these limitations through prospective studies, utilization of higher-quality CXR data, and further refinement of the interpretability methods.