AI Reduces Missed Findings: Results from MGH Study
Massachusetts General Hospital radiologists find that AI helped them miss fewer findings
Chest radiography, also known as a chest X-ray (CXR), is a widely used imaging test for the screening, diagnosis, and monitoring of various cardiothoracic disorders. According to estimates, CXRs make up around 20% of all imaging exams, with millions of CXRs performed in the United States alone each year. Despite its widespread use, CXR interpretation is subjective and prone to wide interobserver inconsistencies. This can lead to missed findings, which can have serious implications, as 19% of early lung cancers that present as nodules on CXRs are missed.
To address this issue, researchers from Massachusetts General Hospital and Harvard Medical School, Qure.ai, and CARPL conducted a study to evaluate the frequency of missed findings in CXRs and the potential of artificial intelligence (AI) to reduce missed findings. The study included 2407 CXRs from eight sites. All CXRs were examined by two thoracic radiologists, who scored each one on a five-point scale for the presence and clinical importance of abnormal findings. The CXRs were also processed with an AI model and the outputs were recorded for the presence of findings.
The results showed that 18.9% (410/2407) of the CXRs had unreported or missed findings, with 76.1% (312/410) of these findings being clinically important. The most common missed findings were pulmonary nodules (38.3%), consolidation (14.6%), linear opacities (9.0%), mediastinal widening (5.1%), hilar enlargement (4.1%), pleural effusions (2.7%), rib fractures (1.4%), and pneumothoraces (0.7%). Following are the examples of missed findings.
The AI model was able to detect 53% (69/131) of the missed findings with an area under the curve (AUC) of up to 0.935. Pneumothorax and mediastinal widening had the lowest AUCs for the AI algorithm, whereas highest AUCs were reported for pleural effusions, enlarged cardiac silhouette, hilar prominence and rib fractures. The AI model was found to be generalizable across different sites, geographic locations, patient genders, and age groups. Examples of findings which were not documented in radiology reports but were detected by AI is below.
These findings have important implications for patient care, as accurately identifying missed findings on CXRs can improve diagnosis and treatment. The use of AI in CXR interpretation may help to improve patient outcomes by providing a second pair of eyes and identifying important missed findings that may have otherwise gone undetected. Further research is needed to evaluate the clinical impact of using AI in CXR interpretation and to determine the optimal use of AI in clinical practice. Overall, AI has the potential to be a valuable tool in the interpretation of CXRs and improving patient care.