Scholarlink

MRI-based age estimation offers a non-invasive approach for assessing biological maturity in forensic medicine. This study compares CNN-,transformer- and hybrid-based architectures for knee MRI age classification and bone segmentation, evaluating vision transformer, Swin transformer, DINOv2, attention U-Net, transformer U-Net, and a baseline CNN U-Net on 1,000 images for classification and 3,000 for segmentation. Pure transformer models failed to train effectively, while attention U-Net and transformer U-Net achieved good segmentation with a dice coefficient of 0.997 but poor classification with an F1-score of 0.751. These results suggest that CNN locality bias is essential for classification, whereas global attention enhances segmentation. Overall, small datasets and the absence of inductive biases limit transformer performance in knee MRI age estimation.

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

Comparison of Modern Transformer Architectures and CNN-based Models for MRI-based Age Estimation of the Knee

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Marco Pawłowski,
  • Dennis Säring,
  • Jochen Herrmann,
  • Eilin Jopp-van Well,
  • Heinz Handels

摘要

MRI-based age estimation offers a non-invasive approach for assessing biological maturity in forensic medicine. This study compares CNN-,transformer- and hybrid-based architectures for knee MRI age classification and bone segmentation, evaluating vision transformer, Swin transformer, DINOv2, attention U-Net, transformer U-Net, and a baseline CNN U-Net on 1,000 images for classification and 3,000 for segmentation. Pure transformer models failed to train effectively, while attention U-Net and transformer U-Net achieved good segmentation with a dice coefficient of 0.997 but poor classification with an F1-score of 0.751. These results suggest that CNN locality bias is essential for classification, whereas global attention enhances segmentation. Overall, small datasets and the absence of inductive biases limit transformer performance in knee MRI age estimation.