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.
Introduction
Diagnosing Alzheimer’s Disease (AD) early is critical for effective intervention, yet traditional methods rely heavily on manual "feature engineering." Radiologists typically identify specific anatomical changes—such as hippocampus shrinkage or ventricle enlargement—which introduces noise, bias, and a heavy reliance on domain expertise. Furthermore, classifiers trained on specific datasets often fail to generalize to new data due to overfitting on dataset-specific quirks.
To address these challenges, this research introduces a 3D Adaptable Convolutional Neural Network (3D-ACNN). This model automatically learns to identify relevant biomarkers directly from MRI scans and is designed to adapt across different medical imaging domains.
Methodology
The 3D-ACNN architecture utilizes a two-stage learning process:
- Unsupervised Pre-training (3D-CAE): A 3D Convolutional Autoencoder is trained to reconstruct 3D MRI volumes. By learning to compress and then reconstruct the entire brain, the network naturally identifies major anatomical structures like the hippocampus and ventricles without explicit labeling.
- Supervised Fine-Tuning: The encoder portion of the pre-trained 3D-CAE serves as the base for a classification network. Fully connected layers are appended and fine-tuned using labeled clinical data for specific diagnostic tasks.
- Domain Adaptation: To handle variations in MRI resolution and dimensions across datasets, the researchers employed "Net2Net" initialization. This allows the model to transfer learned filters to a new target domain while maintaining the integrity of the pre-trained knowledge.
The model utilizes ReLU activation functions and Adadelta for optimization. Notably, the system is robust enough to process "raw" MRI data without the labor-intensive "skull-stripping" preprocessing step.
Experimental Results
The model was pre-trained on the CADDementia dataset and evaluated on a subset of the ADNI (Alzheimer’s Disease Neuroimaging Initiative) dataset, consisting of 210 subjects (70 AD, 70 MCI, 70 NC).
The 3D-ACNN was benchmarked against seven state-of-the-art methods. It demonstrated superior performance across all classification tasks, achieving these results using only MRI data, whereas many competitors required additional PET or CSF biomarkers.
Key performance metrics include:
- AD vs. Normal Control (NC): 97.6% accuracy.
- AD vs. MCI: 95.1% accuracy.
- MCI vs. NC: 90.8% accuracy.
- Ternary Classification (AD/MCI/NC): 89.1% accuracy, a significant improvement over previous multimodal methods (53.8%).
Discussion and Conclusion
Visualization of the network's feature maps confirms that different layers naturally focus on specific AD biomarkers, such as cortical thickness and hippocampus volume. This transparency helps demystify the "black box" of deep learning in medical contexts.
The study demonstrates that features learned from one population are highly transferable to another, suggesting the potential for "foundation models" in neuroimaging. While the current study used a standard sample size of 210 subjects, future work will involve scaling to larger databases and applying this adaptable framework to other pathologies like lung cancer or autism. The 3D-ACNN represents a significant step toward automated, robust, and generalizable diagnostic tools for neurodegenerative diseases.
Summary updated