【摘要】 针对糖尿病患者的分类识别准确率低的问题,提出一种结合静息态功能磁共振成像(MRI)技术和深度学习的方法来完成糖尿病患者和正常对照组的分类识别。首先,对原始的MRI图像进行预处理;再利用低频振幅(ALFF)、局部一致性(Re Ho)方法对预处理后的MRI图像进行特征映射;最后,将特征映射后的图像作为卷积神经网络(CNN)的输入,在卷积神经网络中进行分类识别。实验结果表明,Re Ho特征映射后的MRI图像作为卷积神经网络的输入,正常对照组与糖尿病患者的分类识别正确率为91. 42%,Ⅰ型糖尿病患者与Ⅱ型糖尿病患者的分类识别正确率为94. 82%,正常对照组、Ⅰ型糖尿病患者、Ⅱ型糖尿病患者的分类识别正确率为93. 69%;以ALFF特征映射后的MRI图像作为卷积神经网络的输入,正常对照组与糖尿病患者的分类识别正确率为90. 04%,Ⅰ型糖尿病患者与Ⅱ型糖尿病患者的分类识别正确率为96. 23%,正常对照组、Ⅰ型糖尿病患者、Ⅱ型糖尿病患者的分类识别正确率为94. 48%。分类准确率高于采用模糊支持向量机集成学习。更多还原

【Abstract】 In order to solve the problem of low classification accuracy of diabetic patients, a new method combining resting state functional Magnetic Resonance Imaging( MRI) and deep learning was proposed to accomplish the classification and identification of diabetic patients and normal control group. Firstly, the original MRI images were preprocessed, and then Amplitude of Low-Frequency Fluctuation( ALFF), and Reginal Homogeneity( Re Ho) methods were used to map the features of the pre-processed MRI images. Finally, the feature-mapped image was used as the input of a Convolutional Neural Network( CNN) for classification and recognition. The experimental results show that by using Re Ho feature mapped image as the input of the convolutional neural network, the correct recognition rate were 91. 42% for normal control group and diabetic patients,94. 82% for type Ⅰ diabetic patients and type Ⅱ diabetic patients, and 93. 69% for normal control group, type Ⅰ diabetic patients and type Ⅱ diabetic patients; by using ALFF feature mapped image as the input of the convolutional neural network,the correct recognition rate were 90. 04% for normal control group and diabetic patients, 96. 23% for type Ⅰ diabetic patients and type Ⅱ diabetic patients, and 94. 48% for normal control group, type Ⅰ diabetic patients and type Ⅱ diabetic patients.The accuracies of classification are higher than those by using fuzzy support vector machine.更多还原