Scholarlink

The segmentation of gliomas in magnetic resonance imaging (MRI) is a clinical challenge, especially in clinical settings where access to high-end equipment and computing power is limited. In this paper, we propose a two-phase segmentation framework based on transfer learning applied to multimodal 2D MRI (T1-weighted and T2-FLAIR). The first phase focuses on brain/background separation using U-Net, achieving Dice coefficients greater than 0.95 and false negative rates of less than 1.5% after post-processing. In the second phase, tumor segmentation is developed using a factorial design that combines two architectures (U-Net and Attention U-Net), two input resolutions (120 × 120 and 32 × 32), configurations with and without clinical context, and both modalities, generating a total of 16 models. The results show that the inclusion of clinical context and 32 × 32 centered crops systematically improve performance, achieving a Dice score of 0.9322 and an IoU of 0.8730. In addition, weighted multimodal fusion of T1ce and T2-FLAIR, with α parameter adjustment, consistently outperformed the performance of each modality separately.

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

Glioma Segmentation Model Based on 2D Multimodal Magnetic Resonance Imaging Using a U-Net Transfer Learning Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kevin Liñan,
  • Paulo Valerio,
  • Wilfredo Ticona

摘要

The segmentation of gliomas in magnetic resonance imaging (MRI) is a clinical challenge, especially in clinical settings where access to high-end equipment and computing power is limited. In this paper, we propose a two-phase segmentation framework based on transfer learning applied to multimodal 2D MRI (T1-weighted and T2-FLAIR). The first phase focuses on brain/background separation using U-Net, achieving Dice coefficients greater than 0.95 and false negative rates of less than 1.5% after post-processing. In the second phase, tumor segmentation is developed using a factorial design that combines two architectures (U-Net and Attention U-Net), two input resolutions (120 × 120 and 32 × 32), configurations with and without clinical context, and both modalities, generating a total of 16 models. The results show that the inclusion of clinical context and 32 × 32 centered crops systematically improve performance, achieving a Dice score of 0.9322 and an IoU of 0.8730. In addition, weighted multimodal fusion of T1ce and T2-FLAIR, with α parameter adjustment, consistently outperformed the performance of each modality separately.