MCNet:融合多层感知机和卷积的轻量级病变区域分割网络  

MCNet:a lightweight lesion segmentation network integrating multilayer perceptrons and convolutions

作　　者：申华磊 上官国庆 袁成雨 陈艳浩 刘栋[1,2,3] Shen Hualei;Shangguan Guoqing;Yuan Chengyu;Chen Yanhao;Liu Dong(School of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China;Henan Key Laboratory of Educational Artificial Intelligence and Personalized Learning,Henan Normal University,Xinxiang 453007,China;Big Data for Teaching Resources and Educational Quality Evaluation Henan Engineering Laboratory,Henan Normal University,Xinxiang 453007,China)

机构地区：[1]河南师范大学计算机与信息工程学院,河南新乡453007 [2]河南师范大学河南省教育人工智能与个性化学习重点实验室,河南新乡453007 [3]河南师范大学教学资源与教育质量评估大数据河南省工程实验室,河南新乡453007

出　　处：《河南师范大学学报(自然科学版)》2025年第3期96-103,共8页Journal of Henan Normal University(Natural Science Edition)

基　　金：国家自然科学基金(62072160);河南省科技攻关项目(232102211024).

摘　　要：针对现有医学图像分割网络存在计算量大、对硬件资源要求高和推理速度慢等不足,提出一种轻量级快速分割网络MCNet.MCNet采用编码器-解码器架构,使用多层感知机(MLP)和卷积分别提取并融合医学图像的全局特征和局部特征,以减少网络参数量并提高分割精度.在编码阶段使用卷积分支和多层感知机分支分别提取多尺度的局部特征和全局特征.通过跳跃连接融合这些特征并送入解码器.在解码阶段使用注意力门控机制进行特征增强.在BUSI和ISIC2018数据集上进行实验.和当前最优方法相比,MCNet的Dice相似系数和均交并比在BUSI数据集上分别提高0.11%和0.09%、在ISIC2018数据集上分别提高0.64%和0.95%.同时,MCNet显著减少了网络参数量、降低了浮点运算次数并缩短了CPU推理时间.To address the shortcomings of existing medical image segmentation networks,such as high computational demands,significant hardware resource requirements,and slow inference speeds,a lightweight and fast segmentation network named MCNet is proposed.MCNet adopts an encoder-decoder architecture,utilizing both multilayer perceptron(MLP)and convolutions to extract and fuse global and local features of medical images,respectively,thereby reducing network parameters and improving segmentation accuracy.During the encoding stage,convolutional branches and MLP branches are used to extract multi-scale local and global features.These features are fused via skip connections and passed to the decoder.In the decoding stage,an attention gating mechanism is employed to enhance feature representation.Experiments were conducted on the BUSI and ISIC2018 datasets.Compared with state-of-the-art methods,MCNet achieves improvements in Dice similarity coefficient and mean Intersection over Union of 0.11% and 0.09% on the BUSI dataset,and 0.64% and 0.95% on the ISIC2018 dataset,respectively.Additionally,MCNet significantly reduces the number of network parameters,decreases the number of floating-point operations,and shortens CPU inference time.

关 键 词：医学图像分割 深度神经网络 多层感知机(MLP) 轻量级网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]