Scholarlink

Objective <p>To investigate the feasibility of 2D convolutional neural networks (CNNs) in the automatic classification of anterior talofibular ligaments (ATFLs) on MR images.</p> Methods <p>A total of 560 transverse T2-weighted MR images of the ATFL were collected from Center A, and 96 from Center B; manual segmentation of the ATFL was performed. The ATFL segmentation model was trained on YOLO11 and was validated on images from Center B. The dice similarity coefficient&#xa0;(DSC) between manual and automatic segmentation was calculated. A total of 1,103 T2-weighted MR images of the ATFL were further collected from Center C and divided into three groups: normal, partial, and total tear, and ATFL was automatically segmented for all the images. The 2D ResNet model was then trained for ATFL classification. Finally, the segmentation model and classification model were applied to 420 images from Center D. </p> Results <p>The median DSC for the YOLO11 segmentation model was 0.95. For Center D data, the automatic workflow achieved an accuracy of 92.6% (389/420). It showed 95.0% (190/200) sensitivity and 93.6% (206/220) specificity for abnormal ATFL detection, slightly below the junior radiologist’s 97.0% (194/200) sensitivity and 95.5% (210/220) specificity, but the difference did not reach statistical significance (<i>P</i> = 0.22). Automatic classification of the Center D dataset&#xa0;took 3 minutes, compared with&#xa0;manual 14 minutes for the junior radiologist.</p> Conclusion <p>Automatic segmentation and classification of ATFL on MR images based on CNNs are feasible for evaluating ATFL.</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

Automatic Classification of Anterior Talofibular Ligament Based on 2D Convolutional Neural Network

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Feng Li,
  • Xiao-shan Wang,
  • Ting Li,
  • Gang Wu

摘要

Objective

To investigate the feasibility of 2D convolutional neural networks (CNNs) in the automatic classification of anterior talofibular ligaments (ATFLs) on MR images.

Methods

A total of 560 transverse T2-weighted MR images of the ATFL were collected from Center A, and 96 from Center B; manual segmentation of the ATFL was performed. The ATFL segmentation model was trained on YOLO11 and was validated on images from Center B. The dice similarity coefficient (DSC) between manual and automatic segmentation was calculated. A total of 1,103 T2-weighted MR images of the ATFL were further collected from Center C and divided into three groups: normal, partial, and total tear, and ATFL was automatically segmented for all the images. The 2D ResNet model was then trained for ATFL classification. Finally, the segmentation model and classification model were applied to 420 images from Center D.

Results

The median DSC for the YOLO11 segmentation model was 0.95. For Center D data, the automatic workflow achieved an accuracy of 92.6% (389/420). It showed 95.0% (190/200) sensitivity and 93.6% (206/220) specificity for abnormal ATFL detection, slightly below the junior radiologist’s 97.0% (194/200) sensitivity and 95.5% (210/220) specificity, but the difference did not reach statistical significance (P = 0.22). Automatic classification of the Center D dataset took 3 minutes, compared with manual 14 minutes for the junior radiologist.

Conclusion

Automatic segmentation and classification of ATFL on MR images based on CNNs are feasible for evaluating ATFL.