摘要
医学影像分割在疾病形态分析、活检引导和预后评估中具有重要的临床意义。尽管卷积神经网络(CNN)显著提升了分割性能,但其有限的感受野限制了模型捕获全局上下文特征的能力,尤其是在处理边界模糊和形态不规则的目标时表现较差。因此,文章提出在传统U-Net主干结构基础上,引入轻量化Transformer Bottleneck(LTB)模块,通过局部注意力与低秩投影机制有效捕获全局依赖关系。同时,在跨层连接处设计了跳跃门控跨注意力融合模块(Skip-Gated Cross-Attention,SGCA),以实现特征的选择性融合和边界增强。通过在ISIC2016皮肤病变数据集上的实验验证,该模型在Dice系数(0.927)和IoU指标(0.854)上均优于基线U-Net模型,并且保持了极低的参数量,证明了该方法在医学影像分割任务中的轻量性与有效性。
Medical image segmentation has important clinical significance in disease morphology analysis,biopsy guidance and prognosis evaluation.Although convolutional neural network(CNN)significantly improves the segmentation performance,its limited receptive field limits the ability of the model to capture global context features,especially when dealing with objects with fuzzy boundaries and irregular shapes.Therefore,based on the traditional u-net backbone structure,this paper proposes to introduce the lightweight Transformer Bottleneck(LTB)module to effectively capture global dependencies through local attention and low rank projection mechanism.At the same time,a jump gated cross attention fusion module(SGCA)is designed at the cross layer connection to realize the selective fusion of features and boundary enhancement.Through the experimental verification on ISIC2016 skin lesion data set,the model is superior to the baseline u-net model in Dice coefficient(0.927)and IoU index(0.854),and maintains a very low number of parameters,which proves the lightweight and effectiveness of the method in the task of medical image segmentation.
作者
丁辉
DING Hui(School of Computer Science and Engineering,Anhui University of Science&Technology,Huainan,Anhui 232001,China;Anhui Vocational College of Electronics&Information Technology,Bengbu,Anhui 233000,China)
