Early diagnosis of Alzheimer's disease based on brain morphological changes: A comprehensive approach combining voxel-based morphometry and deep learning

TL;DR

This paper presents a new method for early Alzheimer's diagnosis by combining brain imaging data with deep learning to detect structural changes in the brain.

Contribution

The novel approach integrates biologically meaningful features with deep learning for improved early detection of Alzheimer's disease.

Findings

Biologically driven heatmaps from MRI scans showed excellent classification performance in detecting early Alzheimer's symptoms.

Multi-bit representations of structural biomarkers enhance the interpretability and effectiveness of deep learning models.

The method uses T1-weighted MRI scans from ADNI to analyze five key neuroimaging measures across different clinical groups.

Abstract

Deep learning algorithms optimize data by enhancing resolution and suppressing noise associated with biological knowledge. The root issue is that, for example, CNNs learning mathematical patterns from statistical correlations in the data without regard to biological cues whatsoever, and merely apply filters such as max pooling, never grasping what the biological cues they are supposed to investigate are. This blind procedure can indeed be in technical language; however, it does not help to identify meaningful insights into neuroimaging, where interpretability is essential, and such inadequacies pose a grave challenge. In our research, rather than depending on the CNNs and FCNs only for the feature extractions, we have integrated biologically motivated features into voxel-based morphometry as well as deep learning. Our goal is to analyze T1-weighted MRI scans and T2-Flair images to investigate the characteristics of gray matter, white matter, cerebrospinal fluid, and white matter Hyperintensity in patients with mild cognitive impairment (MCI) who lie on the spectrum between normal aging and Alzheimer's disease (AD). So we extracted critical structural features such as white matter Hyperintensity, gray matter volume, white matter volume, cerebrospinal fluid (CSF) volume, and cortical thickness. These are biologically meaningful biomarkers that reflect the neurodegenerative alterations directly. To validate our method, after the detection of biological features, we have converted them into 3-bit, 4-bit, 8-bit, and 16-bit images. These images were used as inputs for both FCN and CNN models to investigate the early symptoms of AD from classified intracranial features.

•We have used the voxel-based morphometry images as inputs to deep learning algorithms that are mentioned in the manuscript, to identify meaningful insights into the neuroimaging procedure in the early detection of AD.•Our research uses T1-weighted MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI), with structural images from subjects across five clinical groups (CN, EMCI, MCI, LMCI, AD). We derived five well-established imaging measures: total White Matter Hyperintensity (WMH), gray matter (GM) volume, white matter (WM) volume, cerebrospinal fluid (CSF) volume, and cortical thickness, from SPM12 and the CAT12 toolbox.•These measures were transformed into multi-bit (3, 4, 8, and 16-bit) heatmap representations, providing dense, visually encoded inputs for convolutional and fully connected neural networks. Our results indicate that these biologically driven heatmaps provide excellent classification performance, illustrating the complementary value of multimodal structural biomarkers.

We have used the voxel-based morphometry images as inputs to deep learning algorithms that are mentioned in the manuscript, to identify meaningful insights into the neuroimaging procedure in the early detection of AD.

Our research uses T1-weighted MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI), with structural images from subjects across five clinical groups (CN, EMCI, MCI, LMCI, AD). We derived five well-established imaging measures: total White Matter Hyperintensity (WMH), gray matter (GM) volume, white matter (WM) volume, cerebrospinal fluid (CSF) volume, and cortical thickness, from SPM12 and the CAT12 toolbox.

These measures were transformed into multi-bit (3, 4, 8, and 16-bit) heatmap representations, providing dense, visually encoded inputs for convolutional and fully connected neural networks. Our results indicate that these biologically driven heatmaps provide excellent classification performance, illustrating the complementary value of multimodal structural biomarkers.

Figures