Classification of Alzheimer's disease by using tau PET images and deep convolutional neural networks. (December 2021)

• Record Type: Journal Article
• Authors: Ozsahin, Ilker; Sekeroglu, Boran; Butler, Tracy
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 4
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-12
• 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.049148
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Amyloid beta (Ab) and pathologic tau biomarkers are specific neuropathological changes that define Alzheimer's disease (AD). Previously, we applied shallow neural networks for early detection and classification of AD on Ab PET images. The aim of this study is to apply deep convolutional neural network (DCNN) for the classification of AD and mild cognitive impairment (MCI) on tau PET images by using Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Method: Approximately 40, 000 2D images were extracted from 1, 097 baseline and follow‐up tau PET scans (AV‐1451) including 86 AD, 442 MCI and 569 cognitively normal (CN) subjects. All images were fully preprocessed by Laboratory of Neuroimaging (also known as LONI) for making the PET data more uniform and PET images more similar if they are from different scanners. It includes co‐registration, averaging, standardization, and uniform resolution. Scanning protocols are as follows: 370 MBq (10.0 mCi) ± 10% 18 F‐flortaucipir, 30 min (5 min per frame) acquisition at 75‐105 min post‐injection. 2‐class (AD vs CN and MCI vs CN) and 3‐class (AD vs MCI vs CN) experiments were performed. The experiments consisted of image preprocessing and DCNN phases. An image sharpening technique was applied to obtain images with sharpened edges. The histogram equalization and an arithmetic operation were applied to the original images to produce intensity adjusted versions. Then, an image transformation function was applied. Finally, two DCNNs with two convolutional (64x32 filters) and two dense layers (128x8 neurons), and three convolutional (256x64 filters) and three dense layers (128x64x8 neurons) have been trained for 2‐class and 3‐class experiments, respectively. Five‐fold cross‐validation was used. Result: In the 2‐class experiments, receiver operating characteristics area under curve (ROC AUC) scores were found as follows: 0.9943 for AD vs CN and 0.9908 for MCI vs CN. In the 3‐class experiments, macro‐averaged F1 score was found as 0.9827. All results are presented in Tables 1‐4. Figure 1 shows sample Grad‐CAM images. Conclusion: DCNN architecture was accurate in distinguishing AD, MCI, and CN subjects based solely on tau PET images. Further analyses considering Ab status and other clinical factors, and inspection of Grad‐CAM images to understand the neuroanatomic basis of prediction, are underway.