Quality assessment of radiographs through AI-driven feedback: A randomized comparative study on radiographer response in lateral knee radiographs.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology and Nuclear Medicine, University Hospital of Southern Denmark, Esbjerg, Denmark. Electronic address: [email protected].
- Research and Innovation Unit of Radiology, University Hospital of Southern Denmark, Odense, Denmark; Department of Radiology, Odense University Hospital, Odense, Denmark; Centre for Artificial Intelligence (CAI-X), Odense University Hospital, Denmark.
- Radiography Education, VIA University College, Herning, Denmark.
- Department of Radiology and Nuclear Medicine, University Hospital of Southern Denmark, Esbjerg, Denmark.
Abstract
High-quality radiographic positioning is essential for accurate diagnosis in knee imaging. Although artificial intelligence (AI) can assess image quality with high accuracy, its influence on radiographers' clinical decision-making remains unclear. This study examined whether AI feedback affects radiographers' acceptance or rejection of lateral knee radiographs. Eighty-nine participants at European Congress of Radiology 2025 individually assessed 50 lateral knee radiographs (including six repeated cases), with or without AI-generated accept/reject annotations. A pretrained convolutional neural network classified femoral condyle alignment using an expert consensus reference standard. Diagnostic performance, inter-reader agreement (Gwet's AC), and intra-reader reliability (Cohen's kappa) were compared between groups. Radiographers' sensitivity was identical with and without AI support (0.64), with no significant differences in specificity, PPV, or NPV. Inter-reader agreement was lower when AI annotations were shown (AC 0.233 vs. 0.374). Participants with prior AI experience showed improved agreement with AI, whereas agreement decreased among those without experience. False AI feedback reduced sensitivity and specificity in repeated cases, indicating susceptibility to incorrect AI advice. AI feedback did not improve overall performance in assessing lateral knee positioning and may reduce consistency, particularly among users without AI experience. Radiographers demonstrated strong independent assessment skills, highlighting the importance of critical training when implementing AI in radiographic quality assurance. These findings suggest that AI-based quality feedback should be implemented cautiously in radiographic practice, as it may unintentionally reduce consistency among radiographers, particularly those without prior AI experience. Targeted training in critical appraisal of AI outputs and staged integration into clinical workflows are essential to mitigate automation bias.