Abstract
Alzheimer’s disease (AD) remains a significant global health challenge, requiring early and accurate diagnosis to improve patient outcomes. While deep learning has advanced medical imaging analysis, traditional 3D models are often computationally expensive and slow to train. This paper presents a novel broad-deep ensemble model that combines 3D residual convolution for feature extraction with a Broad Learning System (BLS) for efficient classification. Utilizing the ADNI dataset, the model achieved over 92% accuracy in AD detection while reducing training time from nearly an hour to under two seconds.
Introduction
The diagnosis of Alzheimer's disease is often hindered by a shortage of specialized clinicians and the subtle nature of early brain changes. Artificial intelligence, particularly deep learning, has emerged as a vital tool for assisting clinicians. However, standard 3D Convolutional Neural Networks (CNNs) require significant hardware resources and long training cycles. To address these limitations, we propose a hybrid architecture that leverages the spatial feature extraction capabilities of 3D ResNet and the rapid learning speed of the Broad Learning System (BLS).
Methods: The Deep-Broad Architecture
The proposed model splits the diagnostic task into two phases: deep feature extraction and broad classification.
1. Deep Feature Extraction
The model utilizes a 3D Residual Convolution module based on the ResNet architecture. Unlike 2D models that process individual slices, this 3D module analyzes the entire MRI volume, preserving critical spatial relationships between brain structures. This ensures that the complex, three-dimensional atrophy patterns characteristic of AD are captured effectively.
2. Broad Learning System (BLS)
Instead of a traditional deep classifier, features are fed into a BLS. The BLS is a "flat" network that uses feature mapping and enhancement layers. A key advantage of BLS is its weight calculation method; rather than using iterative gradient descent, it employs ridge regression and pseudo-inverse coordinates to determine weights in a single step. This drastically reduces computational overhead and hardware dependency.
Datasets and Preprocessing
The study used data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), specifically the ADNI1 Complete 2Yr 3T standardized dataset. The cohort included 434 subjects (77 AD, 206 MCI, and 151 NC).
Images were preprocessed by resampling to 1.5-mm isotropic voxels and cropping to a 224×224×128 grid. This focused the model's attention on the brain tissue while removing background noise.
Results
The ensemble model was benchmarked against several state-of-the-art architectures, including 3D-ResNet and VoxCNN.
- Accuracy: The model achieved 92.65% accuracy for AD vs. Normal Control and 93.58% for AD vs. MCI. Even the challenging MCI vs. Normal Control task reached 84.68% accuracy.
- Efficiency: While traditional models like VoxCNN required approximately 2,800 seconds to train, the proposed model completed training in just 1.3 to 1.6 seconds.
- Clinical Validation: In a head-to-head comparison with a board-certified nuclear medicine physician, the model significantly outperformed human diagnosis (92% vs. 68.57% accuracy on the same test set), particularly in sensitivity.
Discussion and Conclusion
The deep-broad ensemble model demonstrates that high-performance medical AI does not always require "deeper" architectures. By combining 3D feature extraction with a broad classification layer, we achieve clinical-grade accuracy with unprecedented speed. This efficiency lowers the barrier for real-world clinical deployment, especially in resource-constrained environments. Future work will focus on expanding the dataset size and integrating explainable AI (XAI) to help clinicians interpret the model's findings.
Summary updated