The COVID-19 pandemic has strained healthcare systems globally. Rapid and accurate diagnosis is crucial, but reverse transcription polymerase chain reaction (RT-PCR) testing faces challenges such as delays and limited sensitivity. Chest CT is emerging as a valuable diagnostic tool, exhibiting characteristic patterns in COVID-19 pneumonia. However, visual interpretation can be subjective and time-consuming. Artificial intelligence (AI) offers the potential to automate and improve the accuracy of COVID-19 detection on CT scans. Previous single-center studies have shown promise, but limited generalizability due to data homogeneity. This study aims to develop and evaluate an AI algorithm for COVID-19 detection on chest CT using a large, multinational dataset, aiming to overcome limitations of previous studies by using a globally diverse dataset to improve the generalizability of the AI model.

Existing literature highlights the role of chest CT in diagnosing COVID-19 pneumonia, showing characteristic findings like ground-glass opacities and consolidations. While CT and RT-PCR are often concordant, CT can detect early cases with negative RT-PCR results. However, distinguishing COVID-19 pneumonia from other pneumonias remains a challenge. Several single-center studies explored AI for COVID-19 detection on CT, demonstrating feasibility but limited generalizability. The need for a robust, generalizable AI model trained on a diverse dataset is evident.

This study utilized a large, multinational dataset of chest CT scans from four international centers (Hubei, China; Milan, Italy; Tokyo, Japan; Syracuse, NY, USA). The dataset included 1280 patients with COVID-19 (confirmed by RT-PCR) and a control group of patients with various clinical indications for chest CT, including other pneumonias, cancer diagnoses, and emergency cases. Two deep learning models were developed: a full 3D model and a hybrid 3D model. A lung segmentation algorithm was used to localize the lung regions in the CT scans. Data augmentation techniques were employed to prevent overfitting. Model performance was evaluated using metrics such as accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC). Generalizability was assessed by testing the models on an independent set of patients from Tokyo, Japan, who were excluded from the training and validation sets. Grad-CAM visualizations were used to examine the regions of the CT scans that contributed most to the AI's classification decisions.

The 3D classification model achieved the highest test accuracy (90.8%), sensitivity (84%), and specificity (93%) in an independent test set of 1337 patients. The AUC was 0.949. The hybrid 3D model demonstrated slightly lower performance. The false positive rate in patients with other pneumonias was 10%. When the Tokyo, Japan cohort was removed from training and used for independent testing, the 3D model still maintained acceptable performance, demonstrating good generalizability. Grad-CAM visualizations revealed consistent activation in peripheral lung regions in COVID-19 positive cases, indicating that the AI learned features beyond simple consolidation patterns. The study provides a detailed breakdown of cohort demographics and performance metrics across various patient subgroups.

The findings demonstrate the potential of AI to improve the efficiency and accuracy of COVID-19 pneumonia detection on chest CT scans. The high accuracy and generalizability of the 3D model across diverse patient populations suggest its clinical utility. The high specificity in distinguishing COVID-19 pneumonia from other pneumonias addresses a key challenge in CT interpretation. While CT is not recommended for routine screening or diagnosis, this AI-based tool could assist in settings with limited resources or high prevalence of COVID-19. Future studies should further investigate the algorithm's performance in different clinical settings and explore its potential as a clinical decision support tool.

This study demonstrates the effectiveness of an AI-based approach for detecting COVID-19 pneumonia on chest CT scans using a large, diverse, multinational dataset. The high accuracy, sensitivity, and specificity, along with the demonstrated generalizability to unseen populations, highlight its potential for clinical applications. Future research should focus on integrating this technology into clinical workflows and evaluating its impact on patient outcomes.

The study's limitations include the reliance on RT-PCR confirmation for COVID-19 diagnosis and potential variations in CT scanning protocols and clinical indications across different institutions. The generalizability of the model might be limited to similar patient populations and clinical settings. Further research is needed to validate the model's performance in broader contexts and to refine its capability to differentiate between various types of pneumonia.