Published In

Authors

Abstract

Purpose:
To compare the performance of radiologists to that of a deep learning algorithm for the detection of thoracic abnormalities on chest- X-ray (CXR).

Methods:
Thoracic images of patients who underwent CXR and had a thoracic CT scan within 72 hours at Cochin university hospital were collected over a 10-year period (2010-2020). A senior radiologist specialized in thoracic imaging annotated CXR images for 5 main anomalies (pneumothorax, pleural effusion, mediastino-hilar mass, nodule, and alveolar pattern) and 4 secondary anomalies (cardiomegaly, aortic calcification, hiatal hernia, and foreign bodies), with the corresponding CT scan as the standard of reference. Twelve readers with different levels of expertise were involved in CXR readings, blinded to the expert annotations and CT findings. Their readings were compared to that of a deep learning algorithm (ChestView, Gleamer) that detects thoracic pathologies, trained on 89,229 X-ray images, validated on 3687, and tested on 3722.

Results:
The study included 500 exams of which 266 were abnormal, e. g. with at least one main anomaly seen on CT whereas 234 CXR only showed a secondary anomaly or no anomaly at all. The AI had a sensitivity of 85.0% for the classification of CXR as abnormal and a specificity of 77.4%. The chest radiologists had a sensitivity of 79.9% and specificity of 83.0%, the general radiologists had a sensitivity of 72.2% and a specificity of 82.5%, and the residents had a sensitivity of 73.9% and a specificity of 71.7%. The median reading time was 72s for abnormal CXRs and 48s for normal CXRs.

Conclusion:
These preliminary results show that the AI algorithm has higher sensitivity than all readers for detecting CXR anomalies and higher specificity than junior radiologists.

Limitations:
N/A

Ethics committee approval:
The study received IRB approval (n°AAA-2021-08054).