[Paper Review] Image analysis for Alzheimer's disease prediction: Embracing pathological hallmarks for model architecture design

Sarah C. Brüningk, Felix Hensel+2 more|arXiv (Cornell University)|Nov 12, 2020

Brain Tumor Detection and Classification9 references4 citations

TL;DR

This paper proposes a scalable, hybrid deep learning framework that integrates patch-based 3D convolutional neural networks (3D-CNNs) with topological data analysis (TDA) via persistent homology to simultaneously model local brain atrophy and global connectivity changes in Alzheimer’s disease (AD). The method achieved a mean precision of 0.95±0.03 in classifying cognitively normal individuals and AD patients using T1-weighted MRI data from ADNI, demonstrating competitive performance while enabling interpretability and parallelized training.

ABSTRACT

Alzheimer's disease (AD) is associated with local (e.g. brain tissue atrophy) and global brain changes (loss of cerebral connectivity), which can be detected by high-resolution structural magnetic resonance imaging. Conventionally, these changes and their relation to AD are investigated independently. Here, we introduce a novel, highly-scalable approach that simultaneously captures $ extit{local}$ and $ extit{global}$ changes in the diseased brain. It is based on a neural network architecture that combines patch-based, high-resolution 3D-CNNs with global topological features, evaluating multi-scale brain tissue connectivity. Our local-global approach reached competitive results with an average precision score of $0.95\pm0.03$ for the classification of cognitively normal subjects and AD patients (prevalence $\approx 55\%$).

Motivation & Objective

• To develop a scalable, high-resolution 3D image analysis framework that captures both local and global brain changes in Alzheimer’s disease.
• To integrate topological data analysis (TDA) with deep learning to model multi-scale brain connectivity features robust to noise.
• To overcome computational limitations of full-volume 3D-CNN training by using a patch-based approach with parallelization.
• To improve interpretability of model predictions by linking learned features to pathological hallmarks of AD.
• To establish a foundation for transfer learning to more complex clinical classifications, such as mild cognitive impairment vs. AD.

Proposed method

• A patch-based 3D-CNN architecture was applied to 216 sub-volumes (30×36×30 voxels) extracted from high-resolution T1-weighted MRI scans, with hyperparameters optimized via random search.
• Topological features were extracted using persistent homology on the full 3D MRI volume, with descriptors vectorized via persistence images (PI) for integration into a 2D-CNN.
• Two ensemble strategies were evaluated: (i) combining predictions from all 216 patch-based 3D-CNNs using logistic regression, and (ii) combining a single patch 3D-CNN with global TDA features.
• Image preprocessing included rescaling, cropping, and testing various filters (Gaussian, down-sampling) and persistence filtering to reduce noise.
• The final model used a simple 3D-CNN architecture with He uniform initialization and Adam optimizer, trained for 2500 epochs with early stopping.
• Model interpretability was enabled by analyzing the most predictive image patch and its associated topological features, visualized on sagittal MRI slices.

Experimental results

Research questions

• RQ1 Can a hybrid deep learning model that combines local 3D-CNN features with global topological features improve Alzheimer’s disease classification accuracy?
• RQ2 Does topological data analysis (TDA) with persistent homology enhance the detection of whole-brain connectivity changes in structural MRI?
• RQ3 How does a patch-based 3D-CNN approach compare to full-volume 3D-CNN training in terms of scalability and performance under limited GPU memory?
• RQ4 Can ensemble models that fuse multiple patch predictions and global TDA features achieve higher robustness and generalization than single-model baselines?
• RQ5 To what extent can the model’s predictions be interpreted in terms of known pathological hallmarks of Alzheimer’s disease?

Key findings

• The ensemble model combining all 216 patch-based 3D-CNN predictions achieved a mean precision of 0.95±0.03 in distinguishing cognitively normal subjects from Alzheimer’s disease patients.
• The model using a single representative patch with topological features (PI) showed competitive performance, suggesting that global topological features can complement local patch-level representations.
• Despite extensive preprocessing, topological features did not significantly improve classification performance over the baseline 3D-CNN, indicating potential for refinement in TDA pipeline design.
• The patch-based approach enabled highly parallelized training, achieving inference speeds of 1 second per epoch for 3D-CNNs and 0.1 seconds for 2D-CNNs on standard GPU hardware.
• The method demonstrated strong scalability and interpretability, with the most predictive patch and its associated topological features identifiable and visualizable on MRI slices.
• The study highlights the feasibility of using a simple, interpretable 3D-CNN architecture with TDA for high-resolution 3D MRI analysis, especially when full-volume training is computationally infeasible.