计算机应用研究2026，Vol.43Issue(5)：1585-1593,9.DOI:10.19734/j.issn.1001-3695.2025.07.0291
TriFusion-EdgePVT:融合空间通道多尺度特征与边缘增强的医学图像分割方法
TriFusion-Edge PVT:spatial-channel multi-scale fusion with edge enhancement for medical image segmentation
摘要
Abstract
To address the limitation of convolutional neural network(CNN)in capturing long-range dependencies due to their local receptive fields,this study incorporated a PVTv2 Transformer encoder for powerful global context modeling.Furthermore,to mitigate the Transformer's weaknesses in local context understanding and detail perception,the study designed three novel modules:the residual gated cross-attention module(RGCAM)enhanced activation of relevant features and suppressed irrele-vant information to strengthen both long-range and local contexts;the multi-kernel parallel convolution attention module(MKPCAM)fused spatial-channel attention with multi-scale convolutions to synergistically boost global dependency modeling;the cascaded edge-enhanced upsampling module(CEEUM)hierarchically processed multi-scale edge information to improve detail perception.Experimental results on the Synapse dataset demonstrated that TriFusion-Edge PVT achieved a DSC of 83.61％and HD95 of 15.13.The model also outperformed existing state-of-the-art methods across seven binary medical data-sets.These findings validate the model's comprehensive advantages in global modeling,local semantic preservation,and de-tail recovery,and showcase its cross-dataset generalization capabilities.关键词
医学图像分割/金字塔视觉Transformer/全局上下文建模/多尺度特征融合/注意力机制/细节感知Key words
medical image segmentation/PVT/global context modeling/multi-scale feature fusion/attention mechanism/detail perception分类
信息技术与安全科学引用本文复制引用
刘衢,刘孙俊,王铮,陶奕汀,李刚..TriFusion-EdgePVT:融合空间通道多尺度特征与边缘增强的医学图像分割方法[J].计算机应用研究,2026,43(5):1585-1593,9.基金项目
国家自然科学青年基金资助项目(62101358) （62101358）
四川省科技计划重点研发计划项目(2023YFG0294) （2023YFG0294）
