Abstract
Abnormal β-amyloid (Aβ) accumulation is a primary early indicator of Alzheimer's disease (AD), typically requiring invasive PET scans or cerebrospinal fluid (CSF) assays for detection. This study evaluates the feasibility of predicting Aβ positivity (Aβ+) using non-invasive T1-weighted (T1w) structural MRI. We compared five distinct machine learning and deep learning architectures: classical algorithms (e.g., XGBoost), 2D and 3D Convolutional Neural Networks (CNNs), a hybrid ANN-CNN model, and 3D Vision Transformers. Using data from 1,847 participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI), we found that hybrid models combining 3D image features with anatomical tabular data achieved the highest balanced accuracy (0.815). While performance was lower in cognitively normal individuals, the results suggest that deep learning on standard MRI offers a viable screening tool for patients with mild cognitive impairment (MCI).
Introduction
The landscape of Alzheimer’s disease treatment is shifting toward early intervention with disease-modifying therapies. However, identifying eligible patients remains a bottleneck. The current "gold standard" for detecting amyloid-beta (Aβ) plaques involves expensive PET imaging or invasive lumbar punctures.
There is an urgent clinical need for a "triage" system using widely available, non-invasive tools. While structural MRI does not directly visualize amyloid plaques, the downstream neurodegeneration caused by amyloid buildup creates subtle structural signatures. This paper investigates which computational architectures are most effective at identifying these signatures across the clinical spectrum of healthy aging, MCI, and dementia.
Methodology
Data and Pre-processing
The study utilized a primary dataset of 1,847 participants from the ADNI cohort, including 661 healthy controls, 889 with MCI, and 297 with dementia. Ground truth Aβ status was established via PET Standardized Uptake Value Ratios (SUVR). To enhance model robustness, transfer learning models were pre-trained on a larger dataset of 19,839 subjects from the UK Biobank.
Evaluated Architectures
- Classical Machine Learning: Logistic Regression and XGBoost utilizing tabular features such as age, sex, ApoE4 genotype, and regional brain volumes.
- 2D CNNs: A slice-based approach to process 3D volumes as sequences of 2D images.
- 3D CNNs: Models designed to capture volumetric spatial relationships across the entire brain.
- Hybrid ANN-CNN: An integrated model merging 3D image features with tabular anatomical data (e.g., hippocampal volume).
- 3D Vision Transformers (ViTs): Specifically the Neuroimage Transformer (NiT) and Multiple Instance NiT (MINiT), which utilize self-attention mechanisms to identify long-range spatial dependencies.
Results
The Hybrid Model (3D CNN combined with hippocampal volume) emerged as the top performer, achieving a balanced accuracy of 0.815. Performance varied significantly by clinical subgroup:
- MCI and Dementia: In these symptomatic groups, the 3D CNN reached an accuracy of 0.850.
- Cognitively Normal (CN): Accuracy dropped to 0.631 in healthy controls.
Interestingly, for Vision Transformers, downscaling images to $64^3$ yielded better results than upsampling to $128^3$. This likely reflects the "data-hungry" nature of Transformers; with medical datasets being smaller than standard image sets like ImageNet, lower-resolution inputs may prevent overfitting.
Discussion
The "Jack Model" and Structural Lag
The discrepancy in accuracy between healthy controls and symptomatic patients aligns with the "Jack model" of AD progression. This model posits that amyloid accumulation occurs years before visible structural atrophy appears on an MRI. Consequently, in cognitively normal individuals who are Aβ+, the brain may not yet exhibit the structural changes necessary for an MRI-based model to detect.
The Value of Hybridization
While deep learning models are capable of automated feature extraction, "helping" the model with pre-calculated anatomical data significantly boosts performance. Specifically, providing the volume of the hippocampus and entorhinal cortex—areas typically affected early in AD—allowed the Hybrid ANN-CNN to outperform pure image-based models.
Conclusion
Deep learning models trained on standard T1-weighted MRI show significant promise as a low-cost, non-invasive screening tool for amyloid positivity, particularly in patients already exhibiting mild cognitive impairment. While structural MRI is a "downstream" marker of the disease, its wide availability makes it an ideal candidate for triaging patients before they undergo more invasive and expensive diagnostic procedures. Future work should focus on increasing ancestral diversity in training sets and incorporating white matter microstructural data through Diffusion Tensor Imaging (DTI).
Summary updated