基于改进卷积神经网络的膝关节图像分类研究  

Research on knee joint image classification based on improved CNN

作　　者：李志敏 邹俊忠[1] 张见[1] 王蓓[1] 陈兰岚[1] LI Zhimin;ZOU Junzhong;ZHANG Jian;WANG Bei;CHEN Lanlan(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237

出　　处：《传感器与微系统》2024年第12期49-53,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61976091)。

摘　　要：提取区域有效信息是膝关节磁共振成像(MRI)诊断的关键。为提取MRI图像有效细节特征,提出一种结合注意力机制和上采样融合的深度学习分类模型。首先,通过改进的通道注意力机制增强有用特征,抑制无关特征;然后,利用上采样连接机制改进特征金字塔网络,弥补上采样过程中高级特征的损失问题,并融合多尺度特征;最后,使用MRPyrNet中的细节池化模块对输出的特征图进行不同尺寸的细致分析,增强模型捕捉低级细节特征的能力。在MRNet数据集上的实验结果表明,与其他膝关节MRI图像分类方法相比,所提方法在综合性能方面更有优势。Extracting effective information from regions is crucial for the diagnosis of knee joint magnetic resonance imaging(MRI).To extract the effective details from MRI images,a deep learning classification model that combines attention mechanisms and upsampling fusion is proposed.Firstly,useful features are enhanced and irrelevant features are suppressed through an improved channel attention mechanism.Then,the feature pyramid network(FPN)is improved through an upsampling connection mechanism,which mitigates the loss of highlevel features during the upsampling process and fuses multiscale features.Finally,the detailed pooling module in MRPyrNet is employed to conduct detailed analysis on the output feature map at different scales,thereby enhancing the ability of model to capture lowlevel detail features.Experimental results on the MRNet dataset demonstrate that the proposed method has more advantages compared to other knee joint MRI image classification methods in terms of comprehensive performance.

关 键 词：膝关节 磁共振成像 深度学习 注意力机制 上采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]