Development and External Validation of an Artificial Intelligence-Based Method for Scalable Chest Radiograph Diagnosis: A Multi-Country Cross-Sectional Study.

Problem: Chest radiography is a crucial tool for diagnosing thoracic disorders, but interpretation errors and a lack of qualified practitioners can cause delays in treatment. Aim: This study aimed to develop a reliable multi-classification artificial intelligence (AI) tool to improve the accuracy and efficiency of chest radiograph diagnosis.

Methods

We developed a convolutional neural network (CNN) capable of distinguishing among 26 thoracic diagnoses. The model was trained and externally validated using 795,055 chest radiographs from 13 datasets across 4 countries.

Results

The CNN model achieved an average area under the curve (AUC) of 0.961 across all 26 diagnoses in the testing set. COVID-19 detection achieved perfect accuracy (AUC 1.000, [95% confidence interval {CI}, 1.000 to 1.000]), while effusion or pleural effusion detection showed the lowest accuracy (AUC 0.8453, [95% CI, 0.8417 to 0.8489]). In external validation, the model demonstrated strong reproducibility and generalizability within the local dataset, achieving an AUC of 0.9634 for lung opacity detection (95% CI, 0.9423 to 0.9702). The CNN outperformed both radiologists and nonradiological physicians, particularly in trans-device image recognition. Even for diseases not specifically trained on, such as aortic dissection, the AI model showed considerable scalability and enhanced diagnostic accuracy for physicians of varying experience levels (all P < 0.05). Additionally, our model exhibited no gender bias (P > 0.05).

Conclusion

The developed AI algorithm, now available as professional web-based software, substantively improves chest radiograph interpretation. This research advances medical imaging and offers substantial diagnostic support in clinical settings.