Scholarlink

<p>Multimodal 3D medical image segmentation remains challenging due to feature misalignment caused by intrinsic differences among imaging modalities. To address this issue, we propose the Semantic-Guided Mixture-of-Experts Network (SG-MoE-Net), a modality-aware framework built upon SwinUNETR. The proposed model incorporates a modality-decoupled Mixture-of-Experts architecture, in which heterogeneous experts are selectively activated to process CT, MRI, and semantic representations, thereby reducing cross-modality interference. Furthermore, a cross-level semantic guidance mechanism is introduced to integrate anatomical priors and promote semantic consistency across modalities during feature learning. Extensive experiments on three public multimodal benchmarks, namely CrossMoDA, MM-WHS, and HaN-Seg, demonstrate that SG-MoE-Net consistently outperforms existing state-of-the-art methods. In particular, the proposed approach achieves a Dice score of 84.3% on the CrossMoDA dataset. On the HaN-Seg benchmark, it attains a Dice score of 86.2%, outperforming existing state-of-the-art methods by 1.0% in Dice and 1.4% in Normalized Surface Dice.</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

Sg-moe-net: semantic-guided mixture of experts with cross-modality attention for 3D medical segmentation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Yucai Qu,
  • Canhui Xu

摘要

Multimodal 3D medical image segmentation remains challenging due to feature misalignment caused by intrinsic differences among imaging modalities. To address this issue, we propose the Semantic-Guided Mixture-of-Experts Network (SG-MoE-Net), a modality-aware framework built upon SwinUNETR. The proposed model incorporates a modality-decoupled Mixture-of-Experts architecture, in which heterogeneous experts are selectively activated to process CT, MRI, and semantic representations, thereby reducing cross-modality interference. Furthermore, a cross-level semantic guidance mechanism is introduced to integrate anatomical priors and promote semantic consistency across modalities during feature learning. Extensive experiments on three public multimodal benchmarks, namely CrossMoDA, MM-WHS, and HaN-Seg, demonstrate that SG-MoE-Net consistently outperforms existing state-of-the-art methods. In particular, the proposed approach achieves a Dice score of 84.3% on the CrossMoDA dataset. On the HaN-Seg benchmark, it attains a Dice score of 86.2%, outperforming existing state-of-the-art methods by 1.0% in Dice and 1.4% in Normalized Surface Dice.