Towards Practical Alzheimer's Disease Diagnosis: A Lightweight and Interpretable Spiking Neural Model

TL;DR

This paper introduces FasterSNN, a lightweight and interpretable hybrid spiking-ANN model for Alzheimer's disease diagnosis using unimodal 3D MRI. By combining Leaky Integrate-and-Fire neurons, region-adaptive convolution, and a four-level multi-scale spiking attention framework, the approach achieves strong diagnostic accuracy with low energy consumption and fast inference. Extensive experiments on ADNI and AIBL demonstrate superior performance and robustness against conventional CNNs, Transformers, and prior SNNs, along with interpretable attention maps aligning with known AD pathology. Limitations include reliance on unimodal data and modest temporal modeling; future work targets multimodal integration, harmonization, and atlas-based interpretability to broaden clinical applicability.

Abstract

Early diagnosis of Alzheimer's Disease (AD), particularly at the mild cognitive impairment stage, is essential for timely intervention. However, this process faces significant barriers, including reliance on subjective assessments and the high cost of advanced imaging techniques. While deep learning offers automated solutions to improve diagnostic accuracy, its widespread adoption remains constrained due to high energy requirements and computational demands, particularly in resource-limited settings. Spiking neural networks (SNNs) provide a promising alternative, as their brain-inspired design is well-suited to model the sparse and event-driven patterns characteristic of neural degeneration in AD. These networks offer the potential for developing interpretable, energy-efficient diagnostic tools. Despite their advantages, existing SNNs often suffer from limited expressiveness and challenges in stable training, which reduce their effectiveness in handling complex medical tasks. To address these shortcomings, we introduce FasterSNN, a hybrid neural architecture that combines biologically inspired Leaky Integrate-and-Fire (LIF) neurons with region-adaptive convolution and multi-scale spiking attention mechanisms. This approach facilitates efficient, sparse processing of 3D MRI data while maintaining high diagnostic accuracy. Experimental results on benchmark datasets reveal that FasterSNN delivers competitive performance with significantly enhanced efficiency and training stability, highlighting its potential for practical application in AD screening. Our source code is available at https://github.com/wuchangw/FasterSNN.

Figures (8)

• Figure 1: Overall architecture of the FasterSNN. It stacks multiple Faster SNN blocks: the first three blocks include Depthwise Separable Convolution, Batch Normalization, Gaussian Error Linear Unit, and SWA modules, while the last block excludes the SWA module. The SWA module integrates two types of attention mechanisms: the LIF Spatial Attention and the LIF Channel Attention.
• Figure 2: Performance analysis of various models on the ADNI (left) and AIBL (right) datasets. The radar charts illustrate models' performance on five metrics: energy, training time, number of parameters, accuracy, and Kappa. Each axis represents one metric, normalized between 0 and 1, with spoke length indicating the relative score. The bubble charts compare models based on energy, parameters, and accuracy. The x-axis shows the number of parameters in millions, the y-axis indicates accuracy, and bubble size reflects energy consumption in joules.
• Figure 3: Classification performance metrics of various models on the ADNI and AIBL datasets. The upper half of the rose plot presents six categorical metrics: accuracy, precision, recall, F1-score, Kappa, and average AUC on the ADNI dataset. The lower half shows the same metrics on the AIBL dataset. All models are clearly labeled and color-coded to facilitate visual comparison.
• Figure 4: Model performance evaluation across ADNI and AIBL datasets. Rows 1 and 3 display the confusion matrices for each fold of the 5-fold cross-validation on the ADNI and AIBL datasets, respectively. Rows 2 and 4 present the precision-recall curves for the three diagnostic classes on the ADNI and AIBL datasets correspondingly.
• Figure 5: Confusion matrices of FasterSNN on the ADNI (left) and AIBL (right) datasets, illustrating model's performance across different classes.