Analog Spiking U-Net integrating CBAM&ViT for medical image segmentation

Area of Science:

• Artificial Intelligence
• Deep Learning
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer a low-power, biologically interpretable alternative to traditional Artificial Neural Networks (ANNs) in deep learning.
• Dense prediction tasks, such as medical image segmentation, are increasingly being explored using SNNs.
• Integrating advanced modules like the Convolutional Block Attention Module (CBAM) into SNNs presents a challenge but holds potential for performance enhancement.

Purpose of the Study:

• To propose a novel modified Spiking U-Net architecture, termed Analog Spiking U-Net (AS U-Net), capable of incorporating the Convolutional Block Attention Module (CBAM).
• To evaluate the effectiveness of the AS U-Net with CBAM for diabetic retinal vessel segmentation, aiming to improve accuracy and reduce information loss.
• To demonstrate the state-of-the-art (SOTA) performance and generalization capabilities of the proposed SNN model across various segmentation and image generation tasks.

Main Methods:

• A novel modification to the conventional Spiking U-Net architecture was developed, adjusting neuron firing positions to create the Analog Spiking U-Net (AS U-Net).
• The Convolutional Block Attention Module (CBAM) was successfully integrated into the AS U-Net, marking the first implementation of CBAM within an SNN framework.
• The AS U-Net model was trained using direct encoding on a merged dataset of diabetic retinal vessel segmentation datasets and further tested on ISBI, ISIC, Synapse, and generative tasks.

Main Results:

• The proposed AS U-Net achieved the highest segmentation accuracy in diabetic retinal vessel segmentation, outperforming other SNN and most ANN-based models.
• The model demonstrated comparable performance to ANN models with the same architecture and achieved SOTA results considering both accuracy and energy efficiency.
• Ablative analysis confirmed the feasibility and effectiveness of CBAM within SNNs, indicating its utility for future hardware deployments.

Conclusions:

• The Analog Spiking U-Net (AS U-Net) represents a significant advancement in applying SNNs to dense prediction tasks, particularly in medical image segmentation.
• The successful integration of CBAM into SNNs enhances segmentation performance and reduces information loss, offering a novel approach for efficient AI.
• The AS U-Net exhibits strong generalization capabilities across diverse segmentation and image generation tasks, highlighting its broad applicability and potential for hardware implementation.