FTransCNN:基于模糊融合的Transformer-CNN不确定性医学图像分割模型  

FTransCNN:Fusing Transformer and CNN Based on Fuzzy Fusion for Uncertain Medical Image Segmentation

作　　者：王海鹏 丁卫平[1] 黄嘉爽 鞠恒荣 曹金鑫 刘传升 WANG Haipeng;DING Weiping;HUANG Jiashuang;JU Hengrong;CAO Jinxin;LIU Chuansheng(School of Information Science and Technology,Nantong University,Nantong 226019,China)

机构地区：[1]南通大学信息科学技术学院,江苏南通226019

出　　处：《小型微型计算机系统》2024年第6期1426-1435,共10页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61976120,62006128,62102199)资助;江苏省自然科学基金项目(BK20191445)资助;江苏省双创博士计划项目((2020)30986)资助;江苏省高等学校自然科学研究重大项目(21KJA510004)资助;江苏省高等学校自然科学研究面上项目(21KJB520018)资助;南通市科技局基础科学研究项目(JC2021122)资助;中国博士后科学基金项目(2022M711716)资助;教育部人文社会科学研究青年基金项目(21YJCZH013)资助;江苏省研究生科研与实践创新计划项目(SJCX22_1615)资助。

摘　　要：准确的医学图像分割对于疾病的诊断和治疗至关重要,将多个模型提取的分割特征进行联合使用时存在异质性、不确定性等问题,针对这一问题,本文提出了一种基于模糊融合策略的Transformer-CNN模型(FTransCNN).首先以CNN和Transformer作为骨干网络,并行提取图片的特征信息;其次使用通道注意力促进Transformer全局关键信息,使用空间注意力增强CNN特征局部细节;接着应用Hadamard乘积对来自两个分支特征之间的细粒度交互进行建模,用Choquet模糊积分抑制融合特征的异质性以及融合特征的不确定性;然后使用Fuzzy Attention Fusion Module(FAFM)分级上采样,有效地捕获低级空间特征和高级语义上下文;最后反卷积得到最终的分割结果.在Chest X-ray数据集、Kvasir-SEG数据集和DRIVE数据集上的实验效果表明,FTransCNN与其它深度分割模型相比,在分割任务上有更好的效果.Accurate medical image segmentation is crucial for disease diagnosis and treatment.There are problems of heterogeneity and uncertainty in the joint use of segmentation features extracted by multiple models.In response to this problem,this paper proposes a Transformer-CNN model(FTransCNN)based on a fuzzy fusion strategy.Firstly,CNN and Transformer are used as the backbone network to extract the feature information of pictures in parallel.Secondly,channel attention is used to promote the global key information of Transformer,and spatial attention is used to enhance the local details of CNN features.Thirdly,the proposed model applies Hadamard product to model fine-grained interactions between features from the two branches,and use Choquet fuzzy integral to suppress heterogeneity and uncertainty in fused features.Fourthly,FTransCNN employs Fuzzy Attention Fusion Module(FAFM)hierarchical upsampling to effectively capture low-level spatial features and high-level semantic context.Finally,the new model obtains the final segmentation result by using the deconvolution.The experimental results on Chest X-ray,Kvasir-SEG and DRIVE datasets show that FTransCNN has better performance on segmentation tasks than the-state-of-the-art deep segmentation models.

关 键 词：CHOQUET模糊积分 模糊融合 医学图像分割 TRANSFORMER CNN 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]