AD Multimodal Biomarker Discovery Based On Convolutional Recurrent Neural Network

Alzheimer’s disease(AD)is a brain degenerative disease commonly seen in the elderly.It is irreversible and currently lacks effective curative drugs.Today,the problem of aging society is becoming more and more prominent.More and more elderly people suffer from AD,which has a great impact on their families.Although no effective cure has yet been found,if AD can be identified with high accuracy in the early stages of the disease,effective intervention measures can be implemented to delay the deterioration of the patient’s condition and improve the patient’s quality of life.Also,accurate early detection of AD is important in developing new drugs significance.Structural and functional neuroimaging techniques are effective methods for diagnosing AD.For example,Magnetic Resonance Imaging(MRI)and Positron Emission Tomography(PET)have been proven to help understand the structural anatomy and functional neurological changes associated with AD.With the rapid development of deep learning methods in recent years,Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN)have continuously made breakthroughs in various fields.More and more deep learning method began to be applied to the early diagnosis research of AD.In this paper,the convolutional recurrent neural networks were employed to process both MRI and PET data to improve the performance in the early diagnosis of AD,and at the same time,to determine the biomarkers related to AD.The main contributions are as follows:(1)Combined with the current research hotspots of AD,a multi-slice CNN ensemble model method is proposed.This method decomposes 3D MRI and PET images into a large number of 2D slices along the sagittal,coronal,and transverse planes.Divides these 2D slices into multiple different group data sets according to the decomposition direction.Each group data set is individually used to build a CNN model.After that,the prediction results of the CNN model in three decomposition planes are ensembled and classified.The experimental results show that the method achieved higher accuracy than the compared method,indicating that the multi-slice CNN ensemble model method has strongclassification performance.(2)Aiming at the problem of spatial information loss caused by 3D image slicing operation,we propose a CNN + RNN convolutional recurrent neural network model ensemble method,which is used to extract features within and between 2D slices for ensemble learning.In this paper,CNN + RNN is called the base learner,and the validation set is used to select the top 3 base learners for the sagittal,coronal,and transverse planes accuracy for AD ensemble classification.The experimental results of single-modality and fusion of the two modalities through MRI and PET show that adding RNN model and multi-modality fusion can help improve the accuracy of AD classification to a certain extent.(3)In order to reduce the subjective factors brought by the preset parameters,i.e.Top3,and to consider that the difference between different fold verification sets in 10-fold cross-validation leads to different choices of base learners,we propose a method of combining convolutional recurrent neural network and Genetic Algorithm(GA).In each cross-validation,it selects the combination of base learners with the best performance on the validation set for AD ensemble classification.Experimental results show that this method not only greatly improves the diagnostic performance of AD,but also that most of the AD-related biomarkers we have found are consistent with existing empirical knowledge,which fully demonstrates the effectiveness of our method.