External validation of a pre-trained hybrid convolutional neural network in radiographers agreement of positioning in lateral knee radiographs

Introduction
Accurate positioning in lateral knee radiographs is essential for diagnostic quality but prone to inter-observer variability. Artificial intelligence (AI) may standardize quality assessment, yet its influence on radiographers’ critical reasoning and decisions is unclear. The purpose of this study was to externally validate a pre-trained hybrid convolutional neural network for assessing femoral condyle alignment and to evaluate its effect on radiographers’ classification performance.
Methods
A previously developed AI model (Xception architecture, area under the curve [AUC] = 0.97) was applied to 400 consecutive weight-bearing lateral knee radiographs. Nine clinical diagnostic radiographers from three different institutions independently classified images as accepted or rejected according to predefined positioning criteria, first without AI support and again after a one-month wash-out period with AI assistance consisting of color-coded feedback (green = accepted, red = rejected). Reader performance was compared with a consensus reference using Chi-square tests, diagnostic accuracy measures, fixed-effects meta-analysis, and intra-/inter-reader intraclass correlation coefficients (ICC).
Results
According to the reference standard, 77.7 % of images were acceptable. The AI alone achieved 78.4 % accuracy (sensitivity 52.3 %, specificity 85.8 %). Across readers, AI support increased accepted classifications from 73.4 % to 77.2 % (P < 0.001) and correct classifications from 84.5 % to 89.8 % (P < 0.001). Sensitivity decreased while specificity increased with the use of AI. Inter-reader agreement improved from ICC 0.52 to 0.59.
Conclusion
AI decision support modestly improved accuracy and specificity but did not override professional judgment and clinical reasoning. Radiographers maintained independent decision-making, demonstrating that experienced clinical practitioners were not overruled by AI despite real-time feedback.