Scholarlink

<p>Accurate assessment of femoral component rotation on postoperative total knee arthroplasty (TKA) lateral radiographs is essential but often subjective, leading to unnecessary retakes. We developed a deep learning system to automatically evaluate positioning adequacy and recommend corrective rotation direction (internal or external) when misaligned. A total of 571 images (280 postoperative radiographs and 291 ray-sum computed tomography [CT] images) from 104 TKA cases were cropped, normalized, augmented, and then labeled as acceptable, internal rotation, or external rotation. Two ImageNet-pretrained backbones—ResNet-50 and ConvNeXt-Tiny—were fine-tuned using focal loss, weighted sampling, and stratified 5-fold cross-validation across 30 random seeds. Fifty-six radiographs (20%) were reserved as an independent external test set. A soft probability ensemble combined the two models (ResNet-50: ConvNeXt-Tiny = 0.45: 0.55). Performance was assessed using balanced accuracy (BA), macro precision/recall/F1, area under the curve (AUC), and seed-wise variability. ConvNeXt-Tiny achieved the best performance (BA: 0.829 ± 0.025; macro F1: 0.835; mean AUC: 0.959) and recalled internal rotation cases at 0.93, with seven total misclassifications. The ensemble improved macro precision to 0.887 without increasing BA. ConvNeXt-Tiny significantly outperformed ResNet-50 (<i>p</i> = 0.00054), with low seed-to-seed variance confirming reproducibility. The proposed convolutional neural network reliably assesses positioning adequacy and indicates the required rotation direction on TKA lateral radiographs. Its simplicity and reproducibility suggest its value as a real-time retake support tool to reduce repeat exposures, lower patient dose, and streamline workflow. Future work will include multicenter data to improve external rotation sensitivity and support clinical deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

​​Deep learning-based support system for alignment classification and correction guidance in postoperative total knee arthroplasty lateral radiographs

  • 期刊论文

  • 发表日期：

    2025年11月25日

查看全文
  • Kazuhiro Ogasawara,
  • Shinya Ohwada,
  • Rie Tachibana,
  • Katsuhiko Ogasawara

摘要

Accurate assessment of femoral component rotation on postoperative total knee arthroplasty (TKA) lateral radiographs is essential but often subjective, leading to unnecessary retakes. We developed a deep learning system to automatically evaluate positioning adequacy and recommend corrective rotation direction (internal or external) when misaligned. A total of 571 images (280 postoperative radiographs and 291 ray-sum computed tomography [CT] images) from 104 TKA cases were cropped, normalized, augmented, and then labeled as acceptable, internal rotation, or external rotation. Two ImageNet-pretrained backbones—ResNet-50 and ConvNeXt-Tiny—were fine-tuned using focal loss, weighted sampling, and stratified 5-fold cross-validation across 30 random seeds. Fifty-six radiographs (20%) were reserved as an independent external test set. A soft probability ensemble combined the two models (ResNet-50: ConvNeXt-Tiny = 0.45: 0.55). Performance was assessed using balanced accuracy (BA), macro precision/recall/F1, area under the curve (AUC), and seed-wise variability. ConvNeXt-Tiny achieved the best performance (BA: 0.829 ± 0.025; macro F1: 0.835; mean AUC: 0.959) and recalled internal rotation cases at 0.93, with seven total misclassifications. The ensemble improved macro precision to 0.887 without increasing BA. ConvNeXt-Tiny significantly outperformed ResNet-50 (p = 0.00054), with low seed-to-seed variance confirming reproducibility. The proposed convolutional neural network reliably assesses positioning adequacy and indicates the required rotation direction on TKA lateral radiographs. Its simplicity and reproducibility suggest its value as a real-time retake support tool to reduce repeat exposures, lower patient dose, and streamline workflow. Future work will include multicenter data to improve external rotation sensitivity and support clinical deployment.