Abstract: To address the limited segmentation accuracy caused by intense background interference and blurry boundaries in medical images, a Synergistic Expert Gating Attention Network for medical image segmentation (SEGA-Net) was proposed. First, a Feature Synergistic Aggregation Module (FSAM) is designed to incorporate a statistical distribution recalibration mechanism. By aligning the mean and variance of hierarchical features, the fused features are mapped into a specific distribution interval, thereby suppressing background noise interference arising from cross-layer fusion.Second, An Area-Expert Fusion Module (AEFM) was designed, in which area attention reduces computational complexity while capturing local contextual dependencies; a cascaded Mixture-of-Experts (MoE) network applies multi-expert soft-weighted fusion to perform position-wise differentiated feature enhancement, improving the model's ability to discriminate fine-grained local texture details. Finally, a Multi-receptive Synergistic Module (MRSM) is introduced to construct a dual-domain dynamic multiplicative filter. By performing joint weight allocation in both the channel and spatial domains, the module performs dynamic multiplicative reweighting across channel and spatial domains to suppress redundant multi-scale responses and extract refined lesion boundaries. Experimental results demonstrate that on the Kvasir-SEG dataset, SEGA-Net achieves an improvement of 0.6 percentage points in mean Intersection over Union (mIoU) and 0.58 percentage points in Dice coefficient compared to the EMCAD (Efficient Multi-scale cascaded fully Convolutional Attention Decoder) algorithm, he HD95 metric is reduced by 22.07 compared to the EMCAD algorithm. Furthermore, in the cross-modality generalization experiment on the abdominal MRI (CHAOS-MRI) dataset, the Dice coefficient of the proposed model improved by 1.73 percentage points compared to the baseline model. These results indicate that the proposed method effectively mitigates interference from strong backgrounds and blurry boundaries, enhancing segmentation performance.

Key words: deep learning, neural network, area attention, Mixture of Experts (MoE), medical image segmentation

摘要： 针对医学图像中背景干扰强、边界模糊导致分割精度受限的问题，提出一种基于协同专家门控注意力的医学图像分割网络（SEGA-Net）。首先，设计特征协同聚合模块，引入统计分布重标定机制，通过对齐层级特征的均值与方差，将融合特征映射至分布区间，抑制跨层融合产生的背景噪声干扰；其次，设计区域专家融合模块（AEFM），以区域注意力降低计算复杂度的同时捕获局部上下文依赖，串联混合专家模型（MoE）通过多专家软加权融合对各空间位置的特征进行差异化增强，提升模型对局部纹理细节的分辨能力；最后，引入多感受野协同模块（MRSM），构建双域动态乘性滤波器，通过在通道域与空间域执行联合权重分配，滤除冗余信息，提取病灶精细边缘。实验结果表明，在Kvasir-SEG数据集上，SEGA-Net的平均交并比（mIoU）和Dice系数较EMCAD（Efficient Multi-scale cascaded fullyConvolutional Attention Decoder）算法分别提高0.60和0.58个百分点，豪斯多夫距离HD95较EMCAD算法降低了22.07。在CHAOS-MRI跨模态泛化实验中，模型Dice系数较对比网络提升了1.73个百分点。可见，所提方法能降低背景与边界模糊干扰，提升分割性能。

关键词: 深度学习, 神经网络, 区域注意力, 混合专家模型, 医学图像分割