Development and validation of a deep learning-based automated Notch Width Index measurement system from anteroposterior knee radiographs.
Authors
Affiliations (4)
Affiliations (4)
- Graduate School of Informatics, Gunma University, Maebashi, Japan. Electronic address: [email protected].
- Graduate School of Informatics, Gunma University, Maebashi, Japan.
- Department of Orthopaedic Surgery, Keijinkai Shiroyama Hospital, Ota, Gunma, Japan; Department of Functional Morphology, Nihon University School of Medicine, Tokyo, Japan.
- Department of Orthopaedic Surgery, Keijinkai Shiroyama Hospital, Ota, Gunma, Japan.
Abstract
The Notch Width Index (NWI), a radiographic marker of anterior cruciate ligament (ACL) injury risk, is poorly reproducible when measured manually. We developed a fully automated NWI system combining deep learning segmentation with geometric analysis. A segmentation model (U-Net++/EfficientNet-B5) and a YOLOv8 laterality classifier were trained on 384 public images. Thirty-three anteroposterior knee radiographs were each measured twice by each of two surgeons. The fully automated method (D) was compared with both surgeons, and two semi-automated methods (B, C) with Surgeon 1 to isolate each processing stage, using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. The models were also applied without retraining to 500 external radiographs to assess feasibility and failure modes. Single-reading manual reliability was poor to moderate (ICC 0.475 and 0.562). The deterministic automated method agreed with both surgeons at ICC 0.525-0.756 (moderate to good); the semi-automated methods performed comparably (ICC 0.734 and 0.753). Segmentation reached a mean Dice coefficient of 0.975 ± 0.003, and processing was about 540 times faster than manual measurement. On the external dataset, segmentation degraded on osteoarthritic and low-quality images, so plausible values did not guarantee correct segmentation. The method achieved moderate-to-good agreement with expert measurement while ensuring deterministic, examiner-independent reproducibility. Because it was never trained on surgeons' readings, this reflects reproduction of the geometric definition rather than overfitting. The principal barrier is segmentation under domain shift, not the algorithm; the method is best regarded as a reproducible standardization tool for routine anteroposterior views.