Scholarlink

<p>Glioma, particularly high-grade variants such as glioblastomas, are among the most aggressive and complex tumors, presenting significant challenges for diagnosis, treatment planning, and prognosis. Precise segmentation of key tumor subregions, including the core tumor, enhancing tumor, and whole tumor regions, is crucial for evaluating disease progression and tailoring targeted therapies. Traditional segmentation methods often encounter difficulties due to the inherent heterogeneity of glioma, imaging noise, and computational constraints. To overcome these challenges, we suggest a new framework that includes a robust preprocessing procedure followed by a 3D U-Net architecture enhanced with squeeze and excitation (SE) blocks. SE blocks recalibrate channel-wise feature responses, enhancing feature representation and mitigating noise, thereby improving overall segmentation accuracy and computational efficiency. Our advanced model, evaluated using the BraTS 2019 dataset, attained cutting-edge Dice scores of 0.88 for the whole tumor, 0.90 for the core tumor, and 0.85 for the enhancing tumor. Building on these segmentation results, we introduce a multi-head attention neural network (MHA-NN) and a stacking regressor model to predict overall survival. Our method not only enhances tumor segmentation but also significantly improves survival prediction capabilities. This end-to-end framework represents a significant advance in both tumor segmentation and survival prediction, marking a major step forward in glioma treatment planning and research.</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

SegSurvNet: SE-U-net-based glioma segmentation and overall survival prediction via MHA-NN and stacking regressor

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Novsheena Rasool,
  • Javaid Iqbal Bhat

摘要

Glioma, particularly high-grade variants such as glioblastomas, are among the most aggressive and complex tumors, presenting significant challenges for diagnosis, treatment planning, and prognosis. Precise segmentation of key tumor subregions, including the core tumor, enhancing tumor, and whole tumor regions, is crucial for evaluating disease progression and tailoring targeted therapies. Traditional segmentation methods often encounter difficulties due to the inherent heterogeneity of glioma, imaging noise, and computational constraints. To overcome these challenges, we suggest a new framework that includes a robust preprocessing procedure followed by a 3D U-Net architecture enhanced with squeeze and excitation (SE) blocks. SE blocks recalibrate channel-wise feature responses, enhancing feature representation and mitigating noise, thereby improving overall segmentation accuracy and computational efficiency. Our advanced model, evaluated using the BraTS 2019 dataset, attained cutting-edge Dice scores of 0.88 for the whole tumor, 0.90 for the core tumor, and 0.85 for the enhancing tumor. Building on these segmentation results, we introduce a multi-head attention neural network (MHA-NN) and a stacking regressor model to predict overall survival. Our method not only enhances tumor segmentation but also significantly improves survival prediction capabilities. This end-to-end framework represents a significant advance in both tumor segmentation and survival prediction, marking a major step forward in glioma treatment planning and research.