AI-assisted detection for chest X-rays (AID-CXR): a multi-reader multi-case study protocol

A chest X-ray (CXR) is the most common imaging investigation performed worldwide. Advances in machine learning and computer vision technologies have led to the development of several artificial intelligence (AI) tools to detect abnormalities on CXRs, which may expand diagnostic support to a wider field of health professionals. There is a paucity of evidence on the impact of AI algorithms in assisting healthcare professionals (other than radiologists) who regularly review CXR images in their daily practice.

To assess the utility of an AI-based CXR interpretation tool in assisting the diagnostic accuracy, speed and confidence of a varied group of healthcare professionals.

The study will be conducted using 500 retrospectively collected inpatient and emergency department CXRs from two UK hospital trusts. Two fellowship-trained thoracic radiologists with at least 5 years of experience will independently review all studies to establish the ground truth reference standard with arbitration from a third senior radiologist in case of disagreement. The Lunit INSIGHT CXR tool (Seoul, Republic of Korea) will be applied and compared against the reference standard. Area under the receiver operating characteristic curve (AUROC) will be calculated for 10 abnormal findings: pulmonary nodules/mass, consolidation, pneumothorax, atelectasis, calcification, cardiomegaly, fibrosis, mediastinal widening, pleural effusion and pneumoperitoneum. Performance testing will be carried out with readers from various clinical professional groups with and without the assistance of Lunit INSIGHT CXR to evaluate the utility of the algorithm in improving reader accuracy (sensitivity, specificity, AUROC), confidence and speed (paired sample t-test). The study is currently ongoing with a planned end date of 31 December 2024.

The study has been approved by the UK Healthcare Research Authority. The use of anonymised retrospective CXRs has been authorised by Oxford University Hospital's information governance teams. The results will be presented at relevant conferences and published in a peer-reviewed journal.

Protocol ID 310995-B (awaiting approval), ClinicalTrials.gov.