Scholarlink

<p>Accurate medical image segmentation plays a crucial role in disease quantification and treatment evaluation. Although U-Net and its deep learning–based hybrid variants have been widely applied to medical image segmentation, their fixed down-sampling strategies and the simple stacking of convolutional layers in the encoder limit the effective extraction of multi-scale features across different depths. This constraint reduces the richness of feature representation and adversely affects segmentation performance. To address these limitations, this paper proposes a medical image segmentation framework with multiple attention mechanisms and multi-scale feature fusion, termed MAMF-Net. Specifically, the encoder adopts a hybrid CNN and Transformer architecture to enhance both global context modeling and local feature representation. An Adaptive Atrous Dilated Convolution (AADC) module is introduced to aggregate contextual information under multiple receptive fields, thereby compensating for feature loss induced by down-sampling. In the decoder, a Low-level Dual-Channel Attention Gating (LDCAG) mechanism is applied to low-level skip connections to strengthen spatial localization and edge-aware representations while suppressing background noise. Furthermore, a Multi-Scale Global Channel Attention (MSGCA) module is employed at higher decoding stages to jointly fuse multi-level skip features with decoder semantics, enhancing global consistency and enabling progressive reconstruction from high-level semantics to pixel-level details. Extensive experiments conducted on the Synapse multi-organ dataset and the ACDC cardiac dataset demonstrate that MAMF-Net achieves superior performance. On the Synapse dataset, the proposed method attains an average Dice score of 82.55% and a Hausdorff Distance of 25.38 mm, while on the ACDC dataset it achieves an average Dice score of 91.09% and a Hausdorff Distance of 1.28 mm. Moreover, the proposed model maintains favorable computational efficiency, requiring only 24.25 GFLOPs and 26.13 GFLOPs on the two datasets, respectively. Compared with state-of-the-art methods, MAMF-Net exhibits notable improvements in boundary delineation, small-object segmentation, and global semantic consistency, validating the effectiveness of the proposed architecture for multi-scale feature fusion and semantic alignment.</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

MAMF-Net: A medical image segmentation network with multiple attention and multi-scale feature fusion

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Xiaoxuan Ma,
  • Sirui Wang

摘要

Accurate medical image segmentation plays a crucial role in disease quantification and treatment evaluation. Although U-Net and its deep learning–based hybrid variants have been widely applied to medical image segmentation, their fixed down-sampling strategies and the simple stacking of convolutional layers in the encoder limit the effective extraction of multi-scale features across different depths. This constraint reduces the richness of feature representation and adversely affects segmentation performance. To address these limitations, this paper proposes a medical image segmentation framework with multiple attention mechanisms and multi-scale feature fusion, termed MAMF-Net. Specifically, the encoder adopts a hybrid CNN and Transformer architecture to enhance both global context modeling and local feature representation. An Adaptive Atrous Dilated Convolution (AADC) module is introduced to aggregate contextual information under multiple receptive fields, thereby compensating for feature loss induced by down-sampling. In the decoder, a Low-level Dual-Channel Attention Gating (LDCAG) mechanism is applied to low-level skip connections to strengthen spatial localization and edge-aware representations while suppressing background noise. Furthermore, a Multi-Scale Global Channel Attention (MSGCA) module is employed at higher decoding stages to jointly fuse multi-level skip features with decoder semantics, enhancing global consistency and enabling progressive reconstruction from high-level semantics to pixel-level details. Extensive experiments conducted on the Synapse multi-organ dataset and the ACDC cardiac dataset demonstrate that MAMF-Net achieves superior performance. On the Synapse dataset, the proposed method attains an average Dice score of 82.55% and a Hausdorff Distance of 25.38 mm, while on the ACDC dataset it achieves an average Dice score of 91.09% and a Hausdorff Distance of 1.28 mm. Moreover, the proposed model maintains favorable computational efficiency, requiring only 24.25 GFLOPs and 26.13 GFLOPs on the two datasets, respectively. Compared with state-of-the-art methods, MAMF-Net exhibits notable improvements in boundary delineation, small-object segmentation, and global semantic consistency, validating the effectiveness of the proposed architecture for multi-scale feature fusion and semantic alignment.