Seed-net: structure-enhanced encoder-decoder network via dual-attention bridge for 2D medical image segmentation

Abstract

Medical image segmentation (MIS) plays a crucial role in clinical diagnosis and disease prediction. However, existing encoder-decoder segmentation networks still face several structural limitations. First, conventional downsampling operations may discard high-frequency boundary details, resulting in blurred lesion edges. Second, shallow convolutional features are easily affected by local background noise and scale-variable lesion structures, while deep semantic representations often lack effective long-range dependency modeling, limiting global contextual understanding. Third, highly compressed bottleneck features often mix lesion semantics with background artifacts, which enlarges the semantic gap between the encoder and decoder. Finally, standard decoding operations may introduce upsampling artifacts and fail to accurately reconstruct irregular anatomical boundaries. To address these problems, we propose a structure-enhanced encoder-decoder network, termed SEED-Net, for 2D MIS. Specifically, Haar Wavelet Downsampling is introduced to preserve high-frequency structural information during feature compression, while Deep Mamba layers are deployed in deep semantic stages to capture long-range dependencies with linear computational complexity. In the encoder, the proposed encoder gradient multi-scale module introduces a purification-before-interaction strategy, where Grouped Dynamic Gating first recalibrates group-wise structural responses before local-global multi-scale feature interaction, thereby enhancing lesion-relevant structures while suppressing redundant background activations. At the bottleneck, the Large Kernel Dual-Attention Bridge recalibrates spatial and channel responses to isolate lesion-related semantics from background artifacts. In the decoder, the decoder gradient multi-scale Aggregation module improves boundary reconstruction and suppresses upsampling-induced artifacts through multi-receptive feature aggregation. Experimental results on five public datasets, including DSB2018, ISIC 2016, Kvasir-SEG, DRIVE,and LIDC-IDRI, demonstrate the effectiveness of SEED-Net. Specifically, on ISIC 2016, SEED-Net achieves an IoU of 85.54%, Dice of 91.61%, ACC of 95.64%, Spe of 96.25%, and Sen of 93.21%. These results indicate that SEED-Net can effectively preserve fine structural details, reduce background interference, and produce accurate lesion boundaries, showing its potential for reliable MIS.