计算机科学与探索

• 学术研究 •    下一篇

基于双分支并行编码网络的医学图像分割模型

吴灿辉, 王乐, 毛国君, 饶艳莺, 苏宇征   

  1. 1.福建理工大学 计算机科学与数学学院, 福州 350118
    2.福建理工大学 福建省大数据挖掘与应用技术重点实验室, 福州 350118
    3.福建省儿童医院, 福州 350004
    4.福建省妇幼保健院, 福州 350001

Medical Image Segmentation Model Based on Dual-Branch Parallel Encoding Network

WU Canhui,  WANG Le,  MAO Guojun,  RAO Yanying,  SU Yuzheng   

  1. 1. College of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China
    2. Fujian Provincial Key Laboratory of Big Data Mining and Applications, Fujian University of Technology, Fuzhou 350118, China
    3. Fujian Children's Hospital, University, Fuzhou 350004, China
    4. Fujian Provincial Maternity and Child Health Hospital, Fuzhou 350001, China

PDF

摘要： 在磁共振(MRI)和计算机断层成像(CT)等医学图像分割任务中，人体器官的边缘模糊、小器官难识别以及腹部多器官重叠等问题显著制约了分割精度。然而，主流方法如纯CNN难以建模长距离依赖，而纯Transformer对局部细节捕捉不足且计算开销大。针对上述问题，本文利用全局特征与局部特征之间的互补性，提出双分支编码器DPEncoder，并在此基础上构建了医学图像分割模型 PMCNet。DPEncoder采用双分支并行结构：一个分支基于视觉状态空间模型捕捉图像的全局上下文信息与长距离依赖关系；另一分支则利用卷积网络(CNN)提取精细的局部特征和空间细节，并通过通道多尺度卷积特征融合模块有效地增强了模型的复杂特征的表征能力，很好地实现了全局与局部信息的互补融合。PMCNet基于U型结构，由DPEncoder编码器、对应的解码器以及跳跃连接共同组成，能够实现对MRI或CT切片的高精度分割。实验结果表明，所提出模型在Synapse、ACDC和AMOS2022数据集上的Dice指标分别较基于Mamba的模型Swin-Umamba提高了4.06%、1.82%和2.74%，并且在与其他先进模型的对比中也展示出了显著的优势。

关键词: 医学图像分割, 双分支并行编码器, 多尺度特征, 视觉状态空间, 卷积网络

Abstract: Medical image segmentation in MRI and CT faces challenges such as blurred organ edges, small organ recognition, and overlapping structures. However, mainstream approaches like pure CNNs struggle to model long-range dependencies, while pure Transformers exhibit inadequate capture of local details and incur high computational costs. To enhance segmentation accuracy, DPEncoder(Dual-Branch Parallel Encoder) is proposed, and PMCNet (Parallel Multi-scale Feature Fusion Network)based on it. DPEncoder employs a dual-branch parallel encoder, combining a vision state-space model and CNN-based feature extraction, with a multi-scale convolutional fusion module to integrate global and local information effectively. PMCNet, consisting of DPEncoder and a decoder, achieves high-precision segmentation of MRI/CT slices. Experiments show that PMCNet outperforms the Mamba-based Swin-Umamba by 4.06%, 1.82%, and 2.74% in Dice scores on Synapse, ACDC, and AMOS2022 datasets, respectively, and achieves superior performance over other advanced models.

Key words: Medical image segmentation, Dual-Branch Parallel Encoder, Multi-scale feature maps, Visual State Space, CNN

京ICP备13024263号-1
Copyright © 《计算机科学与探索》编辑部
技术支持：北京玛格泰克科技发展有限公司
总访问量
今日访问
在线人数