摘要
提出一种基于3D-WGMobileNet和迁移学习的网络模型,实现对青年抑郁症不同阶段患者的正确分类。首先,对功能磁共振成像(fMRI)数据进行预处理,并通过局部一致性分析将4D fMRI数据转换为3D,进行降维处理。然后,使用迁移学习方法将阿尔茨海默症的特征迁移到提出的3D-WGMobileNet中。利用动态分组卷积构建卷积核的专家权重矩阵,提高模型的表达能力;利用滑窗分组卷积来压缩模型的参数量,增强模型的计算能力。最后,利用3D-WGMobileNet对抑郁症患者图像进行特征提取和分类。在人类连接组项目数据库上的实验结果表明结合迁移学习、动态分组卷积和滑窗分组卷积的3D-WGMobileNet获得较好的分类效果,其中,抑郁症和健康对照组、轻度抑郁症和健康对照组、轻度抑郁症和中度抑郁症的分类准确率分别达到89.00%、85.15%、87.90%,证明本文方法的可行性和有效性。
A novel network model is proposed based on 3D-WGMobileNet and transfer learning to accurately diagnose the stage of youth depression.After functional magenetic resonance imaging(fMRI) data preprocessing and dimensionality reduction by converting 4-dimensional fMRI data into 3-dimension through regional homogeneity approach,transfer learning method is employed to transfer the characteristics of Alzheimer's disease to the proposed 3D-WGMobileNet.The expert weight matrix of convolutional kernel is constructed using dynamic grouping convolution for improving the expression ability of the model.The sliding window grouping convolution is used to reduce the quantity of model parameters and enhance the computing capability.Finally,3D-WGMobileNet is used for the image feature extraction and classification of youth depression.Experimental results on the dataset of human connectome projects show that the 3D-WGMobileNet incorporating transfer learning,dynamic grouping convolution and sliding window grouping convolution exhibit superior performance in classification,achieving 89.00%,85.15% and 87.90% accuracies in classifying depression and healthy controls,mild depression and healthy controls,mild depression and moderate depression,which verifies the feasibility and effectiveness of the proposed method.
作者
郭朝晖
王瑜
马慧鋆
田恒屹
GUO Zhaohui;WANG Yu;MA Huijun;TIAN Hengyi(School of Computer and Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China)
出处
《中国医学物理学杂志》
CSCD
2024年第4期455-462,共8页
Chinese Journal of Medical Physics
基金
北京市自然科学基金-北京市教育委员会科技计划重点项目(KZ202110011015)
北京工商大学2023年研究生科研能力计划提升项目。
