Poster M4 - An Artificial Intelligence (AI) Method to Identify Alzheimer’s Disease Progression

Introduction: Alzheimer's disease (AD) is a progressive neurodegenerative disorder resulting in brain cell death and brain tissue shrinkage. Many studies have demonstrated the pathophysiological mechanism of AD as a continuum: from normal cognition (NC) to early mild cognitive impairment (EMCI), mild cognitive impairment (MCI), late MCI (LMCI), and finally, AD. However, not all MCI cases progress to AD; some remain stable or even improve over time with appropriate treatments. Therefore, identifying the biomarkers of various MCI stages, especially the EMCI that represents the earliest clinical manifestation of AD, can assist clinicians in implementing disease-modification interventions to slow down AD progression. In recent years, various artificial intelligence (AI) tools, such as CNN, DenseNet-121, CNN-LSTM, LeNet-5, applied to brain imaging data—particularly magnetic resonance imaging (MRI)—were proposed to identify various stages of AD progression. They were primarily developed based on various image-based brain features to classify two (NC and AD) or three stages (NC, MCI, and AD) of AD progression. Additionally, most of these brain features, such as surface area, volume of gray and white matter etc., were derived by considering the whole brain instead of its various parcels. Consequently, there remains the need for a novel AI tool capable of detecting AD-induced localized brain lesions using a minimal amount of brain features. Therefore, this study aimed to develop an AI-ensembled network using important brain features and brain regions in order to identify five stages of AD progression—NC, EMCI, MCI, LMCI, and AD—and link their associated brain lesions effectively.

Materials and

Methods: We retrieved T1-weighted structural MRI data of 1000 participants (200 subjects in each of five categories) from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The data preprocessing steps subsequently consisted of brain extraction, bias field correction, image registration, normalization, and brain parcellation. We used the Automated Anatomical Labeling-3 (AAL-3) atlas to parcellate the brain volume of 79x95x79 mm3 into 170 parcels or regions of interest (ROIs). In each ROI, we estimated the volumes of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) to generate 13 brain features as listed in Fig. 1. We analyzed these brain features using the interquartile range method and identified 40 subjects as data outliers. We then divided the dataset of 960 subjects into training, testing, and validation datasets in a 70:15:15 ratio. Our proposed AI-ensembled network was comprised of a random forest (RF) model with 100 decision trees and a multiclass 3D CNN residual network with 18-deep learnable layers (ResNet-18). The RF model was chosen to identify the most important brain features and regions. The ResNet-18 model was trained, tested, and validated under four input combinations: 1) Combination-1: RF-selected brain features and ROIs, 2) Combination-2: original 13 features and 170 ROIs, 3) Combination-3: RF-selected brain features with 170 ROIs, and 4) Combination-4: original 13 features with RF-selected ROIs (Fig. 1). The network hyperparameters were selected once the validation accuracy reached above 60% for all four combinations. We employed a five-fold cross-validation strategy to evaluate the classification performance of the network.

Results, Conclusions, and Discussions:

Results: The RF model respectively identified GM-to-WM, WM-to-CSF, and GM-to-CSF, volumetric ratios as first, second, and third most important features in predicting the AD progression pattern. Out of 60 RF-selected brain regions, ventral posterolateral and pulvinar lateral regions of brain thalamus, left and right amygdala regions, and all hippocampus regions were found as the most affected brain regions across all AD progression stages, indicating that any clinical manifestation (cognitive changes) in these regions during EMCI stage could potentially be used as AD biomarkers. The hyperparameter tuning using the Adam optimizer yielded an optimal epoch size of 50 and batch size of 35 for all input combinations. The tuning process maintained the values of the performance metrics within a range of ±0.05%. Utilizing these hyperparameters, the ResNet-18 network achieved the highest performance metrics (66% F-1 score, 76% sensitivity, and 93.5% specificity) for Combination-1 input condition (three most-important brain features and 60 brain regions). Additionally, the model’s accuracy was further assessed using the five-fold cross-validation for the Combination-1, which resulted in an accuracy of 60.02%.

Discussion: This study developed a novel RF and ResNet-18 ensembled AI network to classify all five major stages (NC, EMCI, MCI, LCMI, and AD) of the disease progression. In comparison to previous AI models in the literature (developed based on ADNI datasets), our model provided higher values across all performance metrics, which could be attributed to our systematic hyperparameter tuning process and the presence of skip connections and residual blocks in the ResNet-18 network. Nonetheless, the performance of this network could be improved by using a larger training dataset in conjunction with a transfer learning function.

Conclusion: The proposed AI-ensembled network is the first-of-its-kind architecture that can effectively identify the AD continuum, particularly the EMCI stage. The implication of such an AI tool in clinical settings can assist potential AD patients in obtaining disease-modifying targeted therapeutic interventions, which can eventually extend their life expectancy.

Acknowledgements (Optional):