Source-linked AI summary
Alzheimer's Disease Diagnostics by a Deeply Supervised Adaptable 3D Convolutional Network
Ehsan Hosseini-Asl, Georgy Gimel'farb, Ayman El-Baz
TL;DR
AD diagnosis requires MRI features that capture anatomical biomarkers while remaining robust across datasets, but existing approaches can be noisy, preprocessing-dependent, or dataset-specific. The paper uses a source-pretrained 3D convolutional autoencoder with adaptable, deeply supervised classification layers, and reports that the resulting DSA-3D-CNN outperforms competing approaches across five task-specific cases.
Problem
Existing sMRI-based AD classifiers can use noisy, preprocessing-dependent, or dataset-specific features, limiting robust diagnosis across datasets.
Method
The DSA-3D-CNN pretrains stacked 3D-CAE layers to learn generic AD biomarkers, then fine-tunes fully connected layers for target-domain classification with deep supervision.
Results
The DSA-3D-CNN outperformed other approaches in all five task-specific cases using ten-fold cross-validation.
Takeaways & Limitations
The framework supports transferable AD feature learning and task-specific classification across different domains.
Takeaways & Limitations
Prior sMRI feature methods are constrained by registration errors, noise, domain-expert preprocessing, multimodal dependence, or dataset bias.
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 proposes 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 \emph{ADNI} MRI dataset with no skull-stripping preprocessing have shown our 3D-CNN outperforms several conventional classifiers by accuracy and robustness. Abilities of the 3D-CNN to generalize the features learnt and adapt to other domains have been validated on the \emph{CADDementia} dataset.
1. INTRODUCTION
The paper motivates an adaptable deep-learning system for early and accurate AD diagnosis from brain MRI. It targets robust feature extraction across datasets while capturing anatomical biomarkers relevant to AD.
- AD causes progressive nerve-cell death and brain-tissue loss, reducing brain volume and impairing functions over time.
- Early, accurate computer-aided diagnosis is critical because AD prevalence and expected care costs are rising.
- The paper develops an adaptable deep-learning system for early AD diagnosis.
- Robust diagnosis should adapt across MRI datasets to reduce distribution discrepancies and group-specific biases.
- Structural MRI is recognized as a promising indicator of AD progression, alongside other investigated neuroimaging modalities.
- The proposed 3D-CNN uses a 3D-CAE to learn generic AD features in lower layers and deep supervision to increase upper-layer specificity.
2. PRIOR WORK
Prior AD classifiers face noisy, preprocessing-dependent, and dataset-specific features. The paper proposes transferable 3D-CNN features learned in one domain and adapted for task-specific classification in another.
- Voxel-wise sMRI features are noisy and require smoothing and clustering before classification.
- Cortical-thickness and hippocampus features can neglect correlated AD-related shape variations elsewhere in the brain, such as ventricle volume.
- Feature vectors depend strongly on registration and noise, making feature engineering reliant on domain expertise.
- Existing trainable feature-extraction and classifier models may depend on multimodal data or remain biased toward their training dataset.
- The proposed 3D-CNN learns AD biomarkers in a source domain and performs task-specific classification in a target domain.
- Its stacked 3D-CAE supplies lower generic-feature layers, while fully connected layers are fine-tuned with deep supervision for target tasks.
3. MODEL
The model combines a hierarchical 3D convolutional autoencoder with adaptable, deeply supervised classification layers. It learns local and hierarchical MRI features, transfers them across domains, and fine-tunes task-specific predictions.
- 3. MODEL: The framework extracts brain-MRI features with a source-domain-trained 3D-CAE and performs target-domain task-specific classification with a deeply supervised 3D-CNN.
- 3.1. 3D Convolutional Autoencoder (3D-CAE): The 3D-CAE uses local convolutional neighborhoods and shared weights to learn image features without the parameter cost of fully connected global mappings.
- 3.1. 3D Convolutional Autoencoder (3D-CAE): Each encoding feature is computed by convolving a fixed 3D voxel neighborhood with learned filters and applying biases and an activation function.
- 3.1. 3D Convolutional Autoencoder (3D-CAE): The autoencoder reconstructs the input image through an inverse transformation using decoding weights, biases, and an activation function.
- 3.1. 3D Convolutional Autoencoder (3D-CAE): Stacked encoding layers pass each layer’s feature image to the next, while max-pooling halves feature-map size and supports higher-level shape variation learning.
- 3.1. 3D Convolutional Autoencoder (3D-CAE): The autoencoder is trained by minimizing mean squared reconstruction error over the training images, with tied decoding and encoding weights reducing free parameters.
- 3.3. Deeply Supervised Adaptive 3D-CNN (DSA-3D-CNN): Transfer learning adapts a source-trained classifier to target data, with generalized lower convolutional features and task-specific fine-tuning in upper layers.
- 3.3. Deeply Supervised Adaptive 3D-CNN (DSA-3D-CNN): Pretraining yields features related to ventricular size, hippocampus shape, and cortical thickness, while Net2Net initialization supports domain-specific kernel and pooling sizes.
4. EXPERIMENTS
Experiments evaluate DSA-3D-CNN feature representations and classification on ADNI, using CADDementia as the source domain and ADNI as the target domain. The learned representations progressively capture anatomical biomarkers and improve class separability, while ten-fold evaluations show robust performance across five classification tasks.
- Experimental setup: The experiments used 30 CADDementia subjects as the source domain and 210 ADNI subjects as the target domain across five classification tasks.The tasks comprised four binary classifications and one ternary AD-versus-MCI-versus-NC classification, evaluated with ten-fold cross-validation.
- Generic and task-specific feature evaluation: The pre-trained 3D-CAE layers captured cortical thickness, brain size, ventricle size, and hippocampus shape across successive feature layers.Lower-layer feature maps were combined to represent increasingly higher-level anatomical variations.
- Generic and task-specific feature evaluation: The generic layers gradually enhanced AD, MCI, and NC discriminability, while task-specific layers further separated the three classes in feature space.The task-specific representation shown for distinguishing AD+MCI from NC placed the class cases at closer distances while maintaining separation.
- Generic and task-specific feature evaluation: The ternary test-set manifold indicated that learned features distinguished AD, MCI, and NC, with subject locations correlating with disease severity.The most severe AD cases appeared at the right-most side of the AD manifold, while the most normal NC cases appeared at the bottom of the NC manifold.
- Classification performance evaluation: The classifier was evaluated with eight metrics, including accuracy, sensitivity, specificity, balanced accuracy, predictive values, F1-score, and ROC AUC.Table 3 summarized means and standard deviations, while Table 4 compared average ten-fold accuracy with seven other approaches.
- Classification performance evaluation: The DSA-3D-CNN outperformed the compared approaches in all five task-specific cases using only sMRI and no prior skull-stripping.The reported comparisons used average results from ten-fold cross-validation.
5. CONCLUSION
The DSA-3D-CNN combines pretrained 3D convolutional autoencoder features with task-specific fully connected layers and deep supervision for AD classification. Evaluations on ADNI, including ROC/AUC and cross-validation comparisons, indicate robust performance and outperformance of state-of-the-art approaches.
- Results: Mean and standard-deviation results across the evaluation metrics indicated high robustness and confidence for AD predictions.The reported performance was summarized for the target ADNI domain across classification tasks.
- Method: Discriminative loss is imposed on the fully connected layers and output classification layers to improve separation among AD, MCI, and NC subjects.The resulting hierarchical features progressively enhance class discriminability in the network's hidden layers.
- Results: ROC curves and AUC scores were evaluated for five ADNI classification tasks, including ternary and binary distinctions among AD, MCI, and NC.The tasks included AD/MCI/NC, AD+MCI/NC, AD/NC, AD/MCI, and MCI/NC.
- Results: Ten-fold cross-validation across seven metrics demonstrated that the proposed DSA-3D-CNN outperformed state-of-the-art models.The evaluation compared the proposed classifier with competing approaches on ADNI data.