Scholarlink

<p>Hybrid CNN–Transformer networks have become effective architectures for 3D medical image segmentation by combining convolutional locality with attention-based context modeling. Most architectural improvements in this direction focus on encoder design and feature fusion, while the decoder hierarchy that reconstructs dense volumetric predictions is often trained mainly through the final segmentation output. In this paper, we ask a complementary question: whether decoder representations in a hybrid 3D segmentation network can be better shaped during training without changing the deployed inference model. We propose DSR (Decoder Supervision and Reconstruction), a training-time decoder strategy for hybrid 3D segmentation networks; we refer to the resulting training scheme and model as DSR-Net. DSR adds auxiliary segmentation heads to intermediate decoder features and a lightweight image reconstruction head to the highest-resolution decoder feature. The segmentation heads provide multi-scale anatomical supervision, while the reconstruction head encourages the final decoder representation to remain image-grounded. All auxiliary branches are removed at inference, so the deployed encoder, decoder, and final segmentation head are unchanged. We evaluate DSR on abdominal CT multi-organ segmentation, vestibular schwannoma MRI segmentation, and prostate MRI zonal segmentation. Across these tasks, DSR achieves the best Dice among the compared methods and shows particularly visible gains in ASD and HD95 where surface-distance metrics are reported. These results suggest that decoder-oriented training can improve hybrid 3D segmentation while leaving the deployed inference model unchanged. The code is publicly available at <a href="https://github.com/HaoLi12345/hybrid_network">https://github.com/HaoLi12345/hybrid_network</a>.</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

DSR-Net: decoder supervision and reconstruction for hybrid 3D medical segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • Hao Li,
  • Nan Jiang

摘要

Hybrid CNN–Transformer networks have become effective architectures for 3D medical image segmentation by combining convolutional locality with attention-based context modeling. Most architectural improvements in this direction focus on encoder design and feature fusion, while the decoder hierarchy that reconstructs dense volumetric predictions is often trained mainly through the final segmentation output. In this paper, we ask a complementary question: whether decoder representations in a hybrid 3D segmentation network can be better shaped during training without changing the deployed inference model. We propose DSR (Decoder Supervision and Reconstruction), a training-time decoder strategy for hybrid 3D segmentation networks; we refer to the resulting training scheme and model as DSR-Net. DSR adds auxiliary segmentation heads to intermediate decoder features and a lightweight image reconstruction head to the highest-resolution decoder feature. The segmentation heads provide multi-scale anatomical supervision, while the reconstruction head encourages the final decoder representation to remain image-grounded. All auxiliary branches are removed at inference, so the deployed encoder, decoder, and final segmentation head are unchanged. We evaluate DSR on abdominal CT multi-organ segmentation, vestibular schwannoma MRI segmentation, and prostate MRI zonal segmentation. Across these tasks, DSR achieves the best Dice among the compared methods and shows particularly visible gains in ASD and HD95 where surface-distance metrics are reported. These results suggest that decoder-oriented training can improve hybrid 3D segmentation while leaving the deployed inference model unchanged. The code is publicly available at https://github.com/HaoLi12345/hybrid_network.