Alzheimer's Disease Detection with a 3D Convolutional Neural Network using Gray Matter Maps from T1‐weighted Brain MRI. (20th December 2022)

• Record Type: Journal Article
• Authors: Dhinagar, Nikhil J; Thomopoulos, Sophia I; Owens‐Walton, Conor; Stripelis, Dimitris; Ambite, Jose Luis; Steeg, Greg Ver; Thompson, Paul M
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 5
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-12-20
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.066446
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: A practical screening tool to detect Alzheimer's disease (AD) based on brain MRI would be valuable. Here we tested a deep learning method for subject‐wise AD classification; as gray matter (GM) is preferentially affected by AD, we also performed MRI tissue classification on the input data, to test the added value of these input features. We set out to compare different types of imaging data types 1 as inputs to a 3D Convolutional Neural Network (CNN) for the AD classification task. Method: We analyzed T1‐weighted brain MRI scans from 1123 subjects (596M/527F, 55.2 ‐ 95.8 years) from the Alzheimer's Disease Neuroimaging Initiative (ADNI). After registering the T1‐w MRI scans to a common brain template, GM was segmented as shown in Figure 1. We subdivided the dataset into training (3, 302 scans/853 subjects), validation (413 scans/100 subjects) and test (170 scans/170 subjects) data for the CNN 2 . We also trained the CNN on different proportions (10%, 20%, 50% and 100%) of the overall ADNI training data. We further validated our results independently on 232 scans/232 subjects (78M/154F, 33‐96 years) from the Open Access Series of Imaging Studies (OASIS) dataset. Result: We summarize our results in Table 1 in terms of the receiver‐operator characteristic curve‐area under the curve (ROC‐AUC). Results were aggregated over two separate runs to demonstrate the model's stability. The CNN with GM segmented T1‐w MRI as input achieved an Average ROC‐AUC of 0.864, compared to 0.859 with the complete T1‐w MRI. The CNN performs better with gray matter segmented input; this performance boost was more pronounced with smaller training sets, i.e., 10% (330 scans) or 20% (660 scans) as seen in Figure 2. The models trained on gray matter maps from ADNI also yielded better performance on the OASIS test set. Conclusion: In this work, we found that using GM extracted from T1‐w MRI scans improves deep learning‐based AD diagnosis. Feature selection is improved by regulating the data input into the CNN. [1] Lu, B., et al., "A Practical Alzheimer's Disease Classifier… on 85, 721 Samples, " bioRxiv Prepr. (2021). [2] Dhinagar, N. J., et al., "3D CNNs for Classification of Alzheimer's … with T1‐Weighted Brain MRI, " SIPAIM (2021).