Objective To determine whether an artificial intelligence candidate could pass the rapid (radiographic) reporting component of the Fellowship of the Royal College of Radiologists (FRCR) examination.
Participants One artificial intelligence candidate (Smarturgences, Milvue) and 26 radiologists who had passed the FRCR examination in the preceding 12 months.
Main outcome measures Accuracy and pass rate of the artificial intelligence compared with radiologists across 10 mock FRCR rapid reporting examinations (each examination containing 30 radiographs, requiring 90% accuracy rate to pass).
Results When non-interpretable images were excluded from the analysis, the artificial intelligence candidate achieved an average overall accuracy of 79.5% (95% confidence interval 74.1% to 84.3%) and passed two of 10 mock FRCR examinations. The average radiologist achieved an average accuracy of 84.8% (76.1-91.9%) and passed four of 10 mock examinations. The sensitivity for the artificial intelligence was 83.6% (95% confidence interval 76.2% to 89.4%) and the specificity was 75.2% (66.7% to 82.5%), compared with summary estimates across all radiologists of 84.1% (81.0% to 87.0%) and 87.3% (85.0% to 89.3%). Across 148/300 radiographs that were correctly interpreted by >90% of radiologists, the artificial intelligence candidate was incorrect in 14/148 (9%). In 20/300 radiographs that most (>50%) radiologists interpreted incorrectly, the artificial intelligence candidate was correct in 10/20 (50%). Most imaging pitfalls related to interpretation of musculoskeletal rather than chest radiographs.
Fig 1Bar charts showing examination percentage scores per Fellowship of the Royal College of Radiologists mock examination, and overall, acquired by artificial intelligence (AI) candidate and radiologist participants in scenario 1 for only “AI interpretable” images (top) and scenario 4 for all images (bottom). Whisker plots denote standard deviation of scores around mean value by all 26 radiologist participantsFig 2Plot of individual sensitivity and false positive rates of 26 radiologists and artificial intelligence (AI), based on scenario 1 (only “AI interpretable” images) considered. Bivariate random effects summary receiver operator characteristic (SROC) curve and summary estimate for radiologists are included for comparison with AI candidateFig 3Normal paediatric abdominal radiograph interpreted by artificial intelligence (AI) candidate as having right basal pneumothorax with dashed bounding box (false positive result). This should have been identified as non-interpretable by AI. French translation: positif=positive; doute=doubt; epanchement pleural=pleural effusion; luxation=dislocation; negatif=negative; nodule pulmonaire=pulmonary nodule; opacite pulmonaire=pulmonary opacificationFig 4Dorsoplantar and oblique views of abnormal right foot radiograph in adult showing osteochondral defect at talar dome (white dashed arrow). This finding was missed by all 26 radiologists and also artificial intelligence candidate (false negative) and was particularly challenging. French translation: doute=doubt; luxation=dislocation; negatif=negative; positif=positiveFig 5Dorsoplantar and oblique views of abnormal right foot radiograph in adult showing acute fracture of proximal phalanx of big toe, correctly interpreted by less than half of radiologists (46%) and correctly identified by artificial intelligence candidate (dashed bounding box). French translation: doute=doubt; luxation=dislocation; negatif=negative; positif=positiveFig 6Abnormal adult pelvic radiograph showing increased sclerosis and expansion of right iliac bone in keeping with Paget’s disease. This was correctly identified by almost all radiologists (96%) but interpreted as normal by artificial intelligence candidate (false negative), given that this was not a pathology it was trained to identify. French translation: doute=doubt; luxation=dislocation; negatif=negative; positif=positiveFig 7Normal lateral scapular Y view of right shoulder in child, incorrectly interpreted by artificial intelligence candidate as having proximal humeral fracture (dashed bounding box). This was false positive result, which was correctly identified as normal by all 26 radiologists. French translation: doute=doubt; epanchement articulaire=joint effusion; negatif=negative; positif=positive
Conclusions When special dispensation for the artificial intelligence candidate was provided (that is, exclusion of non-interpretable images), the artificial intelligence candidate was able to pass two of 10 mock examinations. Potential exists for the artificial intelligence candidate to improve its radiographic interpretation skills by focusing on musculoskeletal cases and learning to interpret radiographs of the axial skeleton and abdomen that are currently considered “non-interpretable.”
