Abstract
INTRODUCTION: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that severely impairs memory and cognition, posing an increasing public health challenge worldwide. Current research suggests that AD is strongly linked to the abnormal accumulation of phosphorylated tau (pTau), a pathological hallmark reflecting neuronal dysfunction and neurodegeneration. Clinically, pTau is typically assessed via cerebrospinal fluid (CSF) analysis or tau positron emission tomography (tau-PET). Although informative, CSF analysis is invasive, and tau-PET requires ionizing radiation. Furthermore, tau-PET is limited in routine clinical settings due to its high cost and radiation exposure. There remains an urgent need for a non-invasive, safe, and widely applicable imaging method to characterize the spatial distribution and concentration of pTau in the human brain to support a biologically based diagnosis of AD.
METHODS: To address this need, this research proposes an imaging framework that integrates multidimensional MRI (MD-MRI) with supervised machine learning to estimate voxelwise pTau concentration and its spatial distribution. The approach leverages the rich microstructural information embedded in the voxelwise diffusion-relaxation probability distributions derived from MD-MRI. Eight postmortem human brain slices from four donors were analyzed. Voxelwise 2D joint distributions of T1, T2, and Mean Diffusivity (T1D and T2D) were derived from MD-MRI, and vectorized joint distributions were applied as input features. Histology-derived pTau concentrations and spatial distributions from the same samples, quantified via immunohistochemistry, served as the ground truth. pTau concentration was stratified into either 2 classes (high vs. low) or 3 classes (low, moderate, high) for subsequent classification. To comprehensively evaluate the framework’s effectiveness, multiple regression and classification models were employed to predict pTau concentration. Before model training, to eliminate the influence of imaging artifacts, only valid voxels within the region of interest were retained according to the binary mask. To simplify the data structure and improve computational efficiency, principal component analysis (PCA) with 95% threshold was applied for dimensionality reduction. During the training phase, a nested cross-validation (5 outer folds and 5 inner folds) was employed to assess model generalizability, while Bayesian optimization was used to select the optimal hyperparameters within each inner loop. For regression tasks, we compared linear regression, quadratic regression, support vector regression (SVR), random forest, and multilayer perceptron (MLP). For 2-class and 3-class classification tasks, we evaluated logistic regression, Fisher’s linear discriminant (FLD), support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP).
RESULTS: The results demonstrated that the random forest model consistently achieved the best and most stable performance across all tasks. For regression, random forest achieved the lowest mean squared errors and highest goodness of fit for both T1D (MSE 0.031 ± 0.001, R2 = 0.797 ± 0.007) and T2D (MSE 0.040 ± 0.001, R2 = 0.724 ± 0.005). For 2-class classification, this model reached accuracies and Cohen’s kappa, respectively, of 0.924±0.002 and 0.803±0.001 for T1D, and 0.909±0.003 and 0.779±0.009 for T2D. For 3-class classification, the model achieved accuracies of 0.879 ± 0.004 for T1D and 0.858 ± 0.005 for T2D, with corresponding Cohen’s kappa values of 0.841 ± 0.004 for T1D and 0.820 ± 0.006 for T2D, respectively. These findings underscore the robustness of the random forest model. It consistently demonstrated superior
predictive capability across both regression and classification tasks. Importantly, the reconstructed imaging maps based on the coordinates of each voxel visually demonstrate a strong concordance between the model predictions and the histology-derived pTau distributions and concentrations, confirming the validity of the approach. Quantitatively, this agreement was confirmed by the structural similarity index measure (SSIM) analysis, with mean SSIM values of 0.82 and 0.81 for regression, 0.90 and 0.89 for binary classification, and 0.87 and 0.86 for 3-class classification on the T1D and T2D data, respectively.
DISCUSSION: These results demonstrate that the T1D and T2D diffusion–relaxation joint distributions of MD-MRI encode rich microstructural information related to tau pathology, and the random forest model can effectively capture this information to predict pTau concentration. The strong spatial correspondence and high SSIM values confirm that the reconstructed images reflect clinically meaningful and biologically relevant patterns of pTau accumulation.
CONCLUSION: Overall, this study establishes a proof of concept that MD-MRI–derived joint distributions can non-invasively capture tau pathology in Alzheimer’s disease. This framework may potentially provide a safe and non-invasive alternative to CSF analysis and tau-PET in clinical diagnosis. In the future, this approach may have potential clinical utility as a marker of tau pathology. Future work will extend this framework to in vivo human datasets to validate its clinical applicability.
Document Type
Article
Class Name
Electrical and Systems Engineering Undergraduate Research
Language
English (en)
Date of Submission
12-5-2025
Recommended Citation
Zhang, Hongyi; Latimer, Caitlin S.; Keene, C. Dirk; Benjamini, Dan; and Kundu, Shinjini, "Mapping Phosphorylated Tau using Multidimensional MRI in Alzheimer’s Disease" (2025). Electrical and Systems Engineering Undergraduate and Graduate Research. 49.
https://openscholarship.wustl.edu/eseundergraduate_research/49