Scholarlink

Moderate-to-Severe Traumatic Brain Injury (msTBI) is a major health concern caused by external forces that result in structural brain damage and can lead to life-threatening conditions. Accurate lesion segmentation remains highly challenging due to the heterogeneity of msTBI, limiting the effectiveness of conventional neuroimaging methods. The AIMS-TBI Segmentation Challenge 2025 was established to advance automated approaches for msTBI lesion segmentation using a large multi-site dataset of T1-weighted MRIs. This dataset included both cases with and without lesions, highlighting the importance of accurate lesion detection. In this work, we present our submission to the challenge, building on the nnU-Net framework and incorporating two enhancements: the augmentation of training data with blended synthetic small lesions and a multiclass strategy to improve small lesion detection. On the unseen test set, our method achieved promising performances, with an overall Dice score of 0.635 across all images.

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

nnU-Net and Synthetic Lesions for Automated Segmentation of Moderate-to-Severe Traumatic Brain Injury

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rachika E. Hamadache,
  • Clara Lisazo,
  • Cansu Yalcin,
  • Valeriia Abramova,
  • Uma M. Lal-Trehan Estrada,
  • Agustin Cartaya Lathulerie,
  • Micaela Rivas Díaz,
  • Adrià Casamitjana,
  • Arnau Oliver,
  • Xavier Lladó

摘要

Moderate-to-Severe Traumatic Brain Injury (msTBI) is a major health concern caused by external forces that result in structural brain damage and can lead to life-threatening conditions. Accurate lesion segmentation remains highly challenging due to the heterogeneity of msTBI, limiting the effectiveness of conventional neuroimaging methods. The AIMS-TBI Segmentation Challenge 2025 was established to advance automated approaches for msTBI lesion segmentation using a large multi-site dataset of T1-weighted MRIs. This dataset included both cases with and without lesions, highlighting the importance of accurate lesion detection. In this work, we present our submission to the challenge, building on the nnU-Net framework and incorporating two enhancements: the augmentation of training data with blended synthetic small lesions and a multiclass strategy to improve small lesion detection. On the unseen test set, our method achieved promising performances, with an overall Dice score of 0.635 across all images.