FasterSNN is a lightweight spiking neural network for Alzheimer's disease diagnosis from 3D MRI scans. It combines bio-inspired LIF neurons with efficient region-adaptive convolutions and multi-scale attention, achieving 89.44% accuracy on ADNI data while using 152× less energy than traditional CNNs. The model's sparse event-driven computation and interpretable architecture make it ideal for clinical deployment.

Figure: FasterSNN Architecture

FasterSNN/
├── README.md
├── requirement.txt
├── main.py
├── train.py
├── test.py
├── util.py
├── model.py
├── dataloader.py
├── plot.py

• Novel Hybrid Architecture: Combines bio-inspired LIF neurons, region-adaptive convolution, and multi-scale spiking attention for efficient AD diagnosis.
• Energy Efficiency: Achieves state-of-the-art accuracy with only 1.15J energy consumption, making it practical for resource-limited settings.
• Clinical Impact: Focuses on early MCI detection (90% accuracy in AIBL dataset) using cost-effective MRI data.
• Open Science: Provides complete source code and pretrained models for reproducibility and community adaptation.
  Figure: Rodar map And Bubble map of FasterSNN on ADNI dataset

FasterSNN integrates three key innovations:

1. LIF Neurons: Mimic biological spiking behavior for sparse, energy-efficient computations.
2. Region-Adaptive Convolution:
   • Core regions: Standard 3D convolution for rich feature extraction.
   • Edge regions: Depthwise separable convolution for efficiency.
3. Multi-Scale Fusion Pyramid:
   • 4-level feature pyramid with learnable cross-layer weights.
   • SWA module dynamically balances channel/spatial attention (α=β=0.5).

Key Equations:

• Membrane potential update: Vₜ = λVₜ₋₁ + (1/T)∑xₜ⁽ⁱ⁾
• Multi-scale fusion: F_fusion = ∑wᵢ⋅Pool(Conv₁ₓ₁(X_out))

Python == 3.10
torch == 2.4.1
nibabel == 5.2.1
numpy == 1.24.4
scikit-learn == 1.3.2
matplotlib == 3.7.5
seaborn == 0.13.2
tqdm == 4.67.1

Organize MRI data in NIfTI format with the following directory structure:

dataset/
├── train/
│ ├── AD/
│ ├── MCI/
│ └── CN/
└── test/
├── AD/
├── MCI/
└── CN/

python train.py \
  --epochs 20 \
  --lr 1e-3 \
  --weight_decay 1e-3 \
  --device cuda:0  # GPU recommended

python test.py \
  --batch_size 16 \
  --device cuda:0

Figure: Confusion matrix of FasterSNN on ADNI dataset

@article{wu2025towards,
  title={Towards Practical Alzheimer's Disease Diagnosis: A Lightweight and Interpretable Spiking Neural Model},
  author={Wu, Changwei and Chen, Yifei and Du, Yuxin and Zong, Jinying and Dong, Jie and Liu, Mingxuan and Peng, Yong and Fan, Jin and Qin, Feiwei and Wang, Changmiao},
  journal={arXiv preprint arXiv:2506.09695},
  year={2025}
}