Fully Connected Multi-Kernel Convolutional Neural Network Based on Alzheimer’s Disease Diagnosis

Article type: Research Article

Authors: Deng, Lan | Wang, Yuanjun^{; *} | Alzheimer’s Disease Neuroimaging Initiative¹

Affiliations: School of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, Shanghai, P.R. China

Correspondence: [*] Correspondence to: Yuanjun Wang, School of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, Shanghai 200093, P.R. China. E-mail: [email protected].

Note: [1] Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: https://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf

Abstract: Background:There is a shortage of clinicians with sufficient expertise in the diagnosis of Alzheimer’s disease (AD), and cerebrospinal fluid biometric collection and positron emission tomography diagnosis are invasive. Therefore, it is of potential significance to obtain high-precision automatic diagnosis results from diffusion tensor imaging (DTI) through deep learning, and simultaneously output feature probability maps to provide clinical auxiliary diagnosis. Objective:We proposed a factorization machine combined neural network (FMCNN) model combining a multi-function convolutional neural network (MCNN) with a fully convolutional network (FCN), while accurately diagnosing AD and mild cognitive impairment (MCI); corresponding fiber bundle visualization results are generated to describe their status. Methods:First, the DTI data is preprocessed to eliminate the influence of external factors. The fiber bundles of the corpus callosum (CC), cingulum (CG), uncinate fasciculus (UNC), and white matter (WM) were then tracked based on deterministic fiber tracking. Then the streamlines are input into CNN, MCNN, and FMCNN as one-dimensional features for classification, and the models are evaluated by performance evaluation indicators. Finally, the fiber risk probability map is output through FMCNN. Results:After comparing the model performance indicators of CNN, MCNN, and FMCNN, it was found that FMCNN showed the best performance in the indicators of accuracy, specificity, sensitivity, and area under the curve. By inputting the fiber bundles of the 10 regions of interest (UNC_L, UNC_R, UNC, CC, CG, CG+UNC, CG+CC, CC+UNC, CG+CC+UNC, and WM into CNN, MCNN, and FMCNN, respectively), WM shows the highest accuracy in CNN, MCNN, and FMCNN, which are 88.41%, 92.07%, and 96.95%, respectively. Conclusion:The FMCNN proposed here can accurately diagnose AD and MCI, and the generated fiber probability map can represent the risk status of AD and MCI.

Keywords: Alzheimer’s disease, deep learning, fiber tracking, fully convolutional networks

DOI: 10.3233/JAD-220519

Journal: Journal of Alzheimer's Disease, vol. 92, no. 1, pp. 209-228, 2023