U-Net-Based Automated Quality Control of Knee Radiographs: Dual-Center Validation and Clinical Intervention

Knee radiography quality directly affects diagnostic accuracy. Current quality control (QC) mainly involves subjective, inefficient manual assessment, while existing automated tools lack sufficient clinical validation. We developed and validated an interpretable artificial intelligence (AI) framework for the automated QC of knee anteroposterior (AP) and lateral (LAT) radiographs and explored its clinical value. This two-stage retrospective study developed and validated U-Net models using 1600 adult single-knee AP and LAT radiographs from 800 patients at two centers. Anatomical segmentation and landmark localization yielded QC indices. Segmentation was evaluated using the Dice similarity coefficient (DSC); landmark localization using Euclidean error, normalized distance error (NDE), and percentage of correct keypoints (PCK); and QC performance using the intraclass correlation coefficient (ICC), sensitivity, and specificity. Separately, six radiographers received 4 weeks of AI-based feedback on 948 radiographs from 474 patients. The mean DSCs were 0.964 and 0.936 in the internal and external validation cohorts, respectively, and most of the ICCs exceeded 0.90. The mean AP localization errors were 13.52 pixels internally and 19.15 pixels externally, and the corresponding LAT values were 26.34 and 41.50 pixels, respectively. PCK@20% was 100.0%/99.4% for AP and 81.6%/61.8% for LAT in the internal/external cohorts. QC sensitivity ranged from 91.67% to 98.36%, and the specificity ranged from 89.13% to 99.10%. Post-feedback sensitivity showed exploratory numerical gains. Effect sizes ranged from 0.30 to 0.73. The proposed framework enables accurate, objective, and automated QC of knee radiographs, as demonstrated through dual-center validation. The workflow evaluation showed favorable exploratory trends in radiographer sensitivity, warranting further evaluation in prospective controlled studies.