The feasibility of double automated reading of chest radiographic screening results (based on the Moscow experiment on computer vision in radiology)

Introduction: Mass preventive chest imaging examinations (fluorography and X-ray) play a critical role in identifying diseases of public health significance, yet place an additional burden on the healthcare system. Automated sorting of imaging studies using medical devices (MD) powered by artificial intelligence (AI) technologies facilitates optimization by relieving radiologists from having to review studies without pathological findings. However, despite high AI accuracy, rare false-negative findings remain possible, which can be critical in medical screening. Objective: The purpose of this retrospective study was to evaluate an automated dual reading method for autonomous sorting of preventive imaging studies. Materials and methods: The study included 411 clinical cases previously misclassified as «normal.» Each study was reanalyzed by a different AI model. Results: As a result, the re-classification yielded 31.6% correct pathology identification and reduced clinically important discrepancies by 25.5%. When simulating the automated dual reading, the proportion of false-negative findings decreased from 0.071% to 0.052%. Discussion: The simulation revealed significant performance variability across the AI models, emphasizing the need for careful post-authorization monitoring to replace underperforming applications. Conclusion: Thus, automated dual reading improves the safety of autonomous sorting by minimizing the number of missed pathological findings. At the same time, the method requires a strategic approach to AI model selection as it poses a risk of reducing the system’s operating efficiency. The optimal benefit-risk ratio should be determined against contribution of preventive programs to public health interest.