3D迁移网络的阿尔茨海默症分类研究

3D Transfer Learning Network for Classification of Alzheimer’s Disease

阿尔兹海默症目前还无法被治愈,若能对其正确诊断,则可采用正确治疗方式延缓病人病情。为减少人工诊断的时间和成本,采用机器学习方法来辅助人工诊断阿尔兹海默症,提出了一种利用3D核磁共振成像信号来诊断的迁移学习方法。该方法采用MobileNet迁移网络来提取瓶颈特征,并增加了一个有监督训练的顶层来进一步降维和提取特征,最后在分类层中将被试者所有切片的特征进行合并和训练,完成阿尔兹海默症与正常控制的分类。该方法的优点在于,可使网络的训练时间下降,提高分类准确率。实验采用了OASIS数据对该方法进行测试,结果表明,该方法的分类准确率比传统迁移学习网络提高了约8个百分点,而时间只有传统迁移方法的1/60。

Alzheimer’s disease cannot be cured at present.If Alzheimer’s disease can be correctly diagnosed,a correct treatment will delay the patient’s condition.To reduce the time and cost of manual diagnosis for Alzheimer’s disease,this paper adopts machine learning to assistantly diagnose Alzheimer’s disease,and proposes a transfer learning method that uses 3D Magnetic Resonance Imaging(MRI)signals.The proposed method uses a separable convolutional network,MobileNet transfer network to complete the classification of Alzheimer’s disease and normal control.The method extracts preliminary features of slices from a raw MRI image on a bottleneck layer,reduces dimensions and extracts further features on a supervised training top layer,and finally combines and trains the features of all slices on a classification layer.The advantage of this transfer learning method to extract image features is that,it can reduce the training time of the network and improve the classification accuracy.In experiment,the proposed method is tested with a group of OASIS data.The experimental results show that compared with the traditional transfer learning network,the classification accuracy of the proposed method can be improved by about 8 percentage points,and the time can be reduced to 1/60.