Scholarlink

Alzheimer’s disease (AD) is characterized by progressive cognitive decline, with hippocampal atrophy serving as a critical neuroimaging biomarker. Accurate segmentation of the hippocampus in MRI scans is essential for quantifying atrophy but remains challenging due to its small size and low contrast. This study proposes an integrated deep learning and radiomics framework to improve AD diagnosis and prognosis. Our novel SRE-Net segmentation architecture introduces three key innovations: a SP-Mamba encoder based on state-space models to capture long-range contextual information, a Spatial-Position Pixel Enhancement (SPPE) module for refined feature extraction, and a Spatial-Channel Attention Redundancy (SCAR) mechanism to suppress irrelevant background information. From the segmented hippocampi, we extract 1,051 3D radiomic features encompassing intensity, shape, and texture characteristics. For diagnostic classification, a bidirectional LSTM (BiLSTM) model achieves exceptional performance (100% accuracy in AD vs. normal controls). Prognostically, our SVM-Random Forest hybrid model predicts MCI-to-AD conversion with an AUC of 0.75. Experimental results demonstrate that SRE-Net significantly outperforms existing methods, achieving a Dice coefficient of 0.898 on the HarP benchmark (2% improvement over CNN/Transformer approaches). This comprehensive pipeline advances both hippocampal segmentation precision and radiomics-based AD assessment, offering a robust tool for clinical decision support and research applications.

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

SR-Net: High-Precision Hippocampal Segmentation and Radiomics-Based Pipeline for Alzheimer’s Disease Diagnosis and Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yifei Chen,
  • Wenzheng Zhao,
  • Yukang Wang,
  • Zhen Wang,
  • Yansen Su,
  • Haitao Li

摘要

Alzheimer’s disease (AD) is characterized by progressive cognitive decline, with hippocampal atrophy serving as a critical neuroimaging biomarker. Accurate segmentation of the hippocampus in MRI scans is essential for quantifying atrophy but remains challenging due to its small size and low contrast. This study proposes an integrated deep learning and radiomics framework to improve AD diagnosis and prognosis. Our novel SRE-Net segmentation architecture introduces three key innovations: a SP-Mamba encoder based on state-space models to capture long-range contextual information, a Spatial-Position Pixel Enhancement (SPPE) module for refined feature extraction, and a Spatial-Channel Attention Redundancy (SCAR) mechanism to suppress irrelevant background information. From the segmented hippocampi, we extract 1,051 3D radiomic features encompassing intensity, shape, and texture characteristics. For diagnostic classification, a bidirectional LSTM (BiLSTM) model achieves exceptional performance (100% accuracy in AD vs. normal controls). Prognostically, our SVM-Random Forest hybrid model predicts MCI-to-AD conversion with an AUC of 0.75. Experimental results demonstrate that SRE-Net significantly outperforms existing methods, achieving a Dice coefficient of 0.898 on the HarP benchmark (2% improvement over CNN/Transformer approaches). This comprehensive pipeline advances both hippocampal segmentation precision and radiomics-based AD assessment, offering a robust tool for clinical decision support and research applications.