Scholarlink

The segmentation of fibroglandular (FGT) and breast tissue in MRI remains an important step for computer assisted radiology and for risk assessment of breast cancer. Currently, all automatic segmentation methods are trained exclusively on T1-based images and segmentations, and there is no existing literature on segmenting non-T1 sequences. To address this research gap, we trained an nnUNet to perform multi-sequence segmentation using solely T1-based masks. Evaluation based on Dice Similarity Scores and visual inspection reveals distinct segmentations per sequence with reasonable overall performance. Our findings indicate that the architecture is able to overcome partially incorrect labels and make use of general breast-specific masks. However, despite the reasonable results, the segmentations missed some potential FGT sites and is likely to improve from sequence-specific masks. Furthermore, the prediction quality is relatively poor for non-dense breasts. Future studies should include reader studies or sequence-specific segmentations as well as a comparative study with and without such sequence-specific segmentations.

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

One for All: UNET Training on Single-Sequence Masks for Multi-sequence Breast MRI Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jarek M. van Dijk,
  • Luyi Han,
  • Luuk Balkenende,
  • Nika Rasoolzadeh,
  • Karine R. Morche,
  • Tianyu Zhang,
  • Ritse M. Mann

摘要

The segmentation of fibroglandular (FGT) and breast tissue in MRI remains an important step for computer assisted radiology and for risk assessment of breast cancer. Currently, all automatic segmentation methods are trained exclusively on T1-based images and segmentations, and there is no existing literature on segmenting non-T1 sequences. To address this research gap, we trained an nnUNet to perform multi-sequence segmentation using solely T1-based masks. Evaluation based on Dice Similarity Scores and visual inspection reveals distinct segmentations per sequence with reasonable overall performance. Our findings indicate that the architecture is able to overcome partially incorrect labels and make use of general breast-specific masks. However, despite the reasonable results, the segmentations missed some potential FGT sites and is likely to improve from sequence-specific masks. Furthermore, the prediction quality is relatively poor for non-dense breasts. Future studies should include reader studies or sequence-specific segmentations as well as a comparative study with and without such sequence-specific segmentations.