Source-linked AI summary
Alzheimer's Disease Diagnostics by Adaptation of 3D Convolutional Network
Ehsan Hosseini-Asl, Robert Keynto, Ayman El-Baz
TL;DR
The paper addresses the need for AD imaging features that capture anatomical biomarkers without relying heavily on handcrafted preprocessing or dataset-specific learning. It proposes a 3D-ACNN that pretrains a 3D convolutional autoencoder and fine-tunes task-specific layers for adaptation across datasets. The classifier outperformed competing approaches across five AD classification tasks using single-modality structural MRI without prior skull stripping.
Problem
Existing AD MRI feature-extraction methods can be noisy, preprocessing-dependent, omit changes outside selected regions, and produce dataset-specific features.
Method
The 3D-ACNN pretrains stacked 3D convolutional autoencoders to extract generic AD-biomarker features, then fine-tunes upper fully connected layers for target-domain classification.
Results
The 3D-ACNN outperformed competing approaches in all five task-specific classification cases using a single sMRI modality without prior skull stripping.
Takeaways & Limitations
Pretraining and layer freezing yielded features that generalized across domains while capturing AD biomarkers in structural brain MRI.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
AD diagnosis requires imaging features that capture disease-related changes across brain structures, but existing approaches can be noisy, preprocessing-dependent, and dataset-specific. The paper proposes a transferable 3D-CNN built on unsupervised 3D-CAE features to address these limitations.
- Motivation: AD progressively reduces brain volume and damages nerve cells and tissue, making early and accurate computer-aided diagnosis important.The projected increase in affected people and care costs further motivates automated diagnosis.
- Existing diagnostic approaches: Structural MRI is a promising indicator of AD progression and supports classification of AD, MCI, and normal-control subjects.Prior systems extract features from structural or functional imaging and train classifiers to distinguish diagnostic groups.
- Existing diagnostic approaches: Existing MRI features include voxel-wise measurements, cortical thickness, and hippocampus shape-volume descriptors.These approaches use structural information at different spatial scales to characterize AD-related changes.
- Limitations of existing features: Voxel-wise features can be noisy and require smoothing or clustering, while cortical and hippocampal features may omit correlated changes in other brain regions.The cited example includes ventricle volume as an affected region not necessarily captured by localized descriptors.
- Limitations of existing features: Feature suitability depends heavily on preprocessing and expert engineering, and many classifiers learn dataset-specific rather than transferable features.Registration errors and noise are identified as contributors to preprocessing dependence.
- Proposed approach: The proposed 3D-ACNN pretrains a 3D convolutional autoencoder for generic feature extraction, then fine-tunes task-specific upper layers across source and target datasets.CADDementia is presented as a source domain and ADNI as a target domain for adaptation.
2. MODEL
The model uses stacked 3D convolutional autoencoders to learn local anatomical features, then adapts task-specific upper layers for Alzheimer’s classification across domains.
- 2. MODEL: The framework encodes brain MRI with a source-domain-trained 3D-CAE and performs task-specific classification using a target-domain-adaptable 3D-CNN.The 3D-CAE provides the feature-extraction component, while the upper network layers are adapted for classification.
- 2. MODEL: The autoencoder reconstructs the original 3D image from encoded feature maps, and training minimizes mean squared reconstruction error over the input images.Encoding and decoding weights are tied by flipping their dimensions, and stochastic gradient descent with back-propagation minimizes the objective.
- 2. MODEL: Stacked convolutional autoencoders use locally connected nodes and shared weights to extract local features from high-dimensional 3D images.This design addresses the computational expense and large training-data requirements of fully connected global feature extraction.
- 2. MODEL: Each 3D feature map is extracted by convolving a fixed n×n×n voxel neighborhood with learned filters and feature-specific biases.The filters operate across the voxel-wise signal components in each neighborhood, followed by a differentiable activation function such as sigmoid or ReLU.
- 2. MODEL: Pre-trained lower convolutional layers extract generic Alzheimer’s-related biomarkers, while fully connected upper layers are fine-tuned for binary or multi-class classification.The reported biomarkers include ventricular size, hippocampus shape, and cortical thickness; Net2Net initialization supports adaptation to target-domain image specifications.
3. EXPERIMENTAL RESULTS
The experiments pre-trained 3D autoencoders on CADDementia and evaluated the adaptable classifier on ADNI across five classification tasks using ten-fold cross-validation.
- 3. EXPERIMENTAL RESULTS: CADDementia structural T1-weighted MRI served as the source domain for pre-training the 3D-CAEs.The scans included probable AD, MCI, and normal-control subjects and were spatially normalized before pre-training.
- 3. EXPERIMENTAL RESULTS: ADNI served as the target domain for five tasks: four binary classifications and one ternary AD-versus-MCI-versus-NC classification.The evaluation used ADNI data without preprocessing or skull stripping, unlike the pre-training setup described for CADDementia.
- 3. EXPERIMENTAL RESULTS: Three hierarchical 3D-CAEs generated progressively lower-dimensional feature maps with dimensions (102 × 76 × 76), (52 × 40 × 40), and (28 × 22 × 22).Each 3D-CAE used eight encoding and decoding filters of size 3 × 3 × 3 and extracted eight feature maps.
- 3. EXPERIMENTAL RESULTS: The 3D-ACNN outperformed competing approaches in accuracy across all five task-specific classification cases.The comparison used ten-fold cross-validation and a single sMRI modality without prior skull stripping.
4. CONCLUSION AND FUTURE WORK
The paper concludes that the 3D-ACNN more accurately predicts Alzheimer’s disease from structural MRI than several state-of-the-art predictors.
- 4. CONCLUSION AND FUTURE WORK: The 3D-ACNN combines three stacked CADDementia-pretrained 3D-CAEs with fully connected layers for AD classification on 210 ADNI subjects.Performance was measured with ten-fold cross-validation and compared with state-of-the-art models.
- 4. CONCLUSION AND FUTURE WORK: The authors identify future applications of the 3D-ACNN in lung cancer, heart failure, and autism detection.These applications are presented as future extensions of the proposed classifier.