Alzheimer's disease diagnostics by adaptation of 3D convolutional network

Ehsan Hosseini-Asl, Robert Keynton, Ayman El-Baz

TL;DR

Researchers have developed a 3D Adaptable Convolutional Neural Network (3D-ACNN) that automatically learns to identify Alzheimer’s Disease (AD) biomarkers from MRI scans. By pre-training a 3D convolutional autoencoder to recognize anatomical structures and then fine-tuning it on clinical data, the model achieves state-of-the-art accuracy (up to 97.6%) without the need for labor-intensive manual feature engineering or skull-stripping.

What Problem Does This Paper Address?

Diagnosing Alzheimer’s Disease early is critical for treatment, but it traditionally relies on "feature engineering." Radiologists and researchers have to manually identify and measure specific brain changes, such as the shrinking of the hippocampus, the thinning of the cortex, or the enlargement of ventricles.

This manual approach has three major flaws:

  1. Noise and Bias: Voxel-wise features are often noisy and highly dependent on specific image preprocessing steps.
  2. Domain Specificity: Classifiers trained on one dataset often fail when applied to another because they are "overfit" to the specific quirks of that data.
  3. Expertise Requirement: It requires significant domain knowledge to decide which anatomical regions matter most.

The authors sought to create a system that could learn these features automatically and, more importantly, transfer that knowledge across different medical datasets.

Methods

The authors propose a two-stage architecture called the 3D Adaptable Convolutional Neural Network (3D-ACNN):

  1. Unsupervised Feature Learning (3D-CAE): They first built a 3D Convolutional Autoencoder. This network is trained to take a 3D MRI volume, compress it into a lower-dimensional "latent space," and then reconstruct the original image. Because it has to reconstruct the whole brain, it naturally learns to recognize major anatomical structures (biomarkers) like the hippocampus and ventricles without being told what they are.
  2. Supervised Fine-Tuning: The "encoder" part of the pre-trained autoencoder is then used as the foundation for a classification network. They added fully connected layers on top and fine-tuned the model using labeled clinical data.
  3. Adaptation (Net2Net): To handle the fact that different MRI datasets might have different resolutions or dimensions, they used "Net2Net" initialization. This allows the network to adapt its learned filters to a new "target domain" (a different dataset) while preserving the knowledge it gained during pre-training.

Notably, the model uses ReLU activation functions and Adadelta for gradient descent, and it performs remarkably well even on "raw" MRI data that hasn't undergone "skull-stripping" (the process of removing non-brain tissue from the image).

Datasets

  • CADDementia: Used as the source domain for pre-training the 3D-CAE. It contains T1-weighted MRI scans of patients with AD, MCI (Mild Cognitive Impairment), and Normal Controls (NC).
  • ADNI (Alzheimer’s Disease Neuroimaging Initiative): Used as the target domain for classification and evaluation. The study used a subset of 210 subjects (70 AD, 70 MCI, 70 NC) to test the model's ability to generalize.

Key Insights and Ideas

  • Visualizing the "Black Box": One of the most compelling parts of the paper is the visualization of the feature maps. The researchers showed that different layers of the 3D-CAE naturally "focus" on specific AD biomarkers: one layer captures cortical thickness, another focuses on brain size, and another specifically isolates the hippocampus and ventricles.
  • Transfer Learning in Medical Imaging: The paper demonstrates that features learned from one population (CADDementia) are highly transferable to another (ADNI). This suggests we can build "foundation models" for neuroimaging that don't need to be retrained from scratch for every new study.
  • Robustness to Preprocessing: Most AD classification models require a "clean" pipeline (registration, skull-stripping, intensity normalization). This 3D-ACNN outperformed competitors even when the skull-stripping step was skipped, suggesting the 3D convolutions are robust enough to ignore "distractor" pixels like the skull or scalp.

Key Results

The 3D-ACNN was tested against seven other state-of-the-art methods (including SVMs and other deep learning models). It outperformed all of them across five different classification tasks:

  • AD vs. Normal Control (NC): 97.6% accuracy (The closest competitor was at 96.8%).
  • AD vs. MCI: 95.1% accuracy.
  • MCI vs. NC: 90.8% accuracy.
  • Ternary Classification (AD/MCI/NC): 89.1% accuracy (Significantly higher than the 53.8% reported by previous multimodal methods).

These results are particularly impressive because the 3D-ACNN achieved them using only MRI data, whereas many of the competing models required a combination of MRI, PET scans, and Cerebrospinal Fluid (CSF) biomarkers.

Limitations and Future Work

  • Dataset Size: While 210 subjects is standard for many neuroimaging papers, it is small by modern deep learning standards. Testing on thousands of subjects from the full ADNI database would further validate the model's robustness.
  • Computational Cost: 3D convolutions are computationally expensive compared to 2D versions, though the authors mitigated this by using a stacked, hierarchical approach.
  • Future Applications: The authors suggest that this "adaptable" 3D-CNN framework isn't limited to Alzheimer's; it could be applied to detect lung cancer, heart failure, or autism by changing the source and target datasets.

Summary updated

Abstract

Early diagnosis, playing an important role in preventing progress and treating the Alzheimer\{'}s disease (AD), is based on classification of features extracted from brain images. The features have to accurately capture main AD-related variations of anatomical brain structures, such as, e.g., ventricles size, hippocampus shape, cortical thickness, and brain volume. This paper proposed to predict the AD with a deep 3D convolutional neural network (3D-CNN), which can learn generic features capturing AD biomarkers and adapt to different domain datasets. The 3D-CNN is built upon a 3D convolutional autoencoder, which is pre-trained to capture anatomical shape variations in structural brain MRI scans. Fully connected upper layers of the 3D-CNN are then fine-tuned for each task-specific AD classification. Experiments on the CADDementia MRI dataset with no skull-stripping preprocessing have shown our 3D-CNN outperforms several conventional classifiers by accuracy. Abilities of the 3D-CNN to generalize the features learnt and adapt to other domains have been validated on the ADNI dataset.
BibTeX
@inproceedings{hosseiniasl2016alzheimer,
  author = {Ehsan Hosseini-Asl and Robert Keynton and Ayman El-Baz},
  title = {{Alzheimer's disease diagnostics by adaptation of 3D convolutional network}},
  booktitle = {ICIP},
  year = {2016},
  url = {https://openreview.net/forum?id=gfySgt593f},
}

Loading related content...

Figure