Diagnosis of normal chest radiographs using an autonomous deep-learning algorithm.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Chest radiography (CXR) is a common diagnostic imaging test.
• Current CXR interpretation can be time-consuming, leading to potential delays and increased workload for radiologists.
• Automated diagnostic tools are being explored to improve efficiency in radiology departments.

Purpose of the Study:

• To assess the effectiveness of a deep learning (DL) algorithm in identifying normal frontal adult chest radiographs (CXRs).
• To evaluate the DL algorithm's suitability as a rule-out test for fully automated diagnosis within an active clinical workflow.

Main Methods:

• A multicentre study involving 3,887 CXRs from four NHS institutions.
• A convolutional neural network (CNN) was utilized to classify examinations with low abnormality scores as high confidence normal (HCN).
• Ground truth (GT) for each radiograph was determined by two independent reviewers, with an arbitrator resolving discrepancies.

Main Results:

• The DL algorithm classified 15% of all CXRs as HCN with 97.7% precision.
• A small proportion (0.33%) of examinations were incorrectly classified as HCN.
• Of the incorrectly classified HCN cases, 84.6% were identified as borderline by the radiologist GT process.

Conclusions:

• Deep learning algorithms demonstrate high precision for automated normality identification in a subset of CXRs.
• Automating the reporting of normal CXRs can significantly reduce radiologist workload and allow focus on complex cases.
• Site-specific algorithm deployment with feedback mechanisms is crucial for optimizing performance and ensuring accurate classification.