DI-MobileNet:基于轻量化网络的骶髂关节炎识别方法  

DI-MobileNet:A lightweight network-based method to identify sacroiliac arthritis

作　　者：杜涛 闫建红[1] DU Tao;YAN Jianhong(School of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,Shanxi,China)

机构地区：[1]太原师范学院计算机科学与技术学院,山西晋中030619

出　　处：《智能计算机与应用》2025年第2期138-143,共6页Intelligent Computer and Applications

基　　金：太原师范学院校级研究生教育创新项目(SYYJSYC-2470);山西省研究生精品教学案例项目(2024AL27)。

摘　　要：针对传统疾病识别网络参数量大,部署效率低的特点,本文提出一种基于改进MobileNetV2模型的分类模型DI-MobileNet,该模型以轻量化网络MobileNetV2为基模型,基于Inception提出多尺度卷积结构;将空洞卷积嵌入模型中,以更高效的方式提取骶髂关节图像不同尺度特征。采用公共数据集Digital Knee X-ray与私有骶髂关节数据集进行实验验证,实验结果表明DI-MobileNet模型参数量为2.23 M,远低于常规卷积神经网络,在Digital Knee X-ray数据集上准确率达到93.05%,在骶髂关节数据集上准确率达到了97.33%,均高于其他模型,较原模型分别提高了1.4和6.31个百分点。Aiming at the characteristics of sacroiliac joint diseases with large intraclass variation and small interclass variation,a classification model DI-MobileNet is proposed to improve the MobileNetV2 model,which takes the lightweight network MobileNetV2 as the base model,proposes a multi-scale convolution structure based on Inception,and subsequently embeds the null convolution into the model to extract sacroiliac iliac joint images with different scale features.The public dataset Digital Knee X-ray with sacroiliac joint dataset is used for validation.Experiments show that the number of parameters of DI-MobileNet model is 2.23 M,which is much lower than that of conventional convolutional neural network.The accuracy reached 93.05%on the Digital Knee X-ray dataset and 97.33%on the sacroiliac joint dataset,which were both higher than the other models,and improved 1.4 and 6.31 percentage points compared with the original model,respectively.

关 键 词：骶髂关节 INCEPTION 空洞卷积 MobileNetV2 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]