Altered structural connectivity detected with dilated convolutional neural network analysis in the DIAN study and the Wisconsin Registry for Alzheimer’s Prevention

• Published in: Alzheimer's & dementia Vol. 17; no. S4
• Main Authors: Zhen, Xingjian, Chakraborty, Rudrasis, Vogt, Nicholas M., Wang, Ruochen, Yang, Kao Lee, Adluru, Nagesh, Gordon, Brian A., Benzinger, Tammie L.S., McKay, Nicole S., Betthauser, Tobey J., Johnson, Sterling C., Singh, Vikas, Bendlin, Barbara B.
• Format: Journal Article
• Language: English
• Published: 01.12.2021
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.054181

Background Dominantly inherited Alzheimer’s disease (DIAD) and late onset Alzheimer’s disease (LOAD) are characterized by the accumulation of amyloid pathology, and neurodegeneration which heralds the onset of dementia. Loss of structural connectivity prior to development of dementia may be measured using techniques that are sensitive to subtle neurodegeneration such as diffusion MRI (dMRI). We have previously shown the utility of deep learning (using a dilated convolutional neural network (DCNN) model) for analysis of sequential manifold‐valued data. Here, we apply this approach to dMRI data to test for loss of structural connectivity among mutation carriers who will develop DIAD, as well as among individuals who are at risk for LOAD due to amyloid pathology. Method Dataset 1 comprised 170 cognitively unimpaired participants from the Wisconsin Registry for Alzheimer’s Prevention study who underwent [11C]PiB‐PET to determine amyloid status, and dataset 2 comprised 440 participants in the Dominantly Inherited Alzheimer Network (DIAN) study. Demographics are shown in Table 2. Participants underwent diffusion weighted imaging which was processed using MRTrix3 and FSL toolkits. TractSeg was performed on both datasets to generate 50 white matter tracts, and tractometry was performed to generate mean representations of the tracts. Similar to previous work, we trained the dilated CNN model (Figure 1) using DTI values along the tracts for each group. The distance between the parameters of two models is treated as the difference between two groups. We performed permutation testing of 5000 runs on the distance to determine significant group differences within each dataset. Result The p‐values for each tract are shown in Table 1. Within the 50 fibers, we identified 14 tracts that differed by amyloid status, and 16 tracts that differed by mutation status. Across the two data sets, 9 tracts, e.g. Arcuate fascicle, and Cingulum, were found to be in common. Conclusion We demonstrate the ability to use the dilated CNN model to capture alterations along tract fibers among cognitively unimpaired individuals with preclinical amyloid as well as among mutation carriers who will develop DIAD. Longitudinal studies are needed to determine the temporal relationship between the accumulation of amyloid and neurodegeneration in the development of dementia.