Scholarlink

<p>Currently, PET/CT imaging technology is widely used in clinical medicine, with one of the main research areas being the automatic segmentation of lesion tissues by combining data from PET/CT scans. However, accurately identifying lesion tissues in PET/CT images remains challenging due to complex tumor boundaries and internal heterogeneity. To address this, we propose MTS-Net, a multi-stage, multi-modal, and multi-task segmentation model. MTS-Net integrates four core components: a Detection Attention Module (DAM) for assisting tumor localization and reducing missed detections, a Deformable Convolution Residual Block (DRconv) to enhance the receptive field during the downsampling phase, a Cross-Fusion Attention Module (CFAM) for multi-scale feature fusion, and an Edge-aware Multi-scale Supervision Module (EAMSM) utilizing deep supervision to refine boundary segmentation. The model effectively combines object detection with segmentation, leveraging cross-modal information and deep edge supervision. Experimental results demonstrate that MTS-Net significantly outperforms existing medical segmentation methods, showing superior accuracy and robustness on the neuroblastoma and HECKTOR2021 datasets.</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

MTS-Net: research on multi-modal, multi-task, multi-stage tumor segmentation model for PET/CT

  • 期刊论文

  • 发表日期：

    2025年7月14日

查看全文
  • Yongwei Zheng,
  • Zhanquan Sun,
  • Suyun Chen,
  • Hongliang Fu,
  • Chaoli Wang,
  • Ji Zhu

摘要

Currently, PET/CT imaging technology is widely used in clinical medicine, with one of the main research areas being the automatic segmentation of lesion tissues by combining data from PET/CT scans. However, accurately identifying lesion tissues in PET/CT images remains challenging due to complex tumor boundaries and internal heterogeneity. To address this, we propose MTS-Net, a multi-stage, multi-modal, and multi-task segmentation model. MTS-Net integrates four core components: a Detection Attention Module (DAM) for assisting tumor localization and reducing missed detections, a Deformable Convolution Residual Block (DRconv) to enhance the receptive field during the downsampling phase, a Cross-Fusion Attention Module (CFAM) for multi-scale feature fusion, and an Edge-aware Multi-scale Supervision Module (EAMSM) utilizing deep supervision to refine boundary segmentation. The model effectively combines object detection with segmentation, leveraging cross-modal information and deep edge supervision. Experimental results demonstrate that MTS-Net significantly outperforms existing medical segmentation methods, showing superior accuracy and robustness on the neuroblastoma and HECKTOR2021 datasets.