Spiking neural networks fine-tuning for brain image segmentation

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Deep learning relies on artificial neural networks (ANNs), which demand significant computational resources.
• Spiking neural networks (SNNs) offer a low-power alternative due to their sparse, event-driven nature.
• Training SNNs is more challenging than training ANNs, hindering their widespread adoption.

Purpose of the Study:

• To develop an efficient and accurate SNN training scheme for segmenting human hippocampi from MRI data.
• To investigate the challenges in ANN-to-SNN conversion and propose a hybrid training strategy.
• To evaluate the impact of data representation (binary/ternary) on SNN performance.

Main Methods:

• A three-stage hybrid training pipeline: ANN optimization, ANN-to-SNN conversion, and spike-based backpropagation fine-tuning.
• Investigated output scaling issues in converted SNNs.
• Empirically evaluated binary and ternary representations in SNNs for image classification and segmentation.

Main Results:

• The proposed hybrid training scheme significantly improved segmentation accuracy and training efficiency over standard ANN-SNN conversion and direct SNN training.
• Experimental results validated the effectiveness of the hybrid approach for hippocampus segmentation.
• The study provided insights into performance decline factors and the influence of data representations.

Conclusions:

• The novel three-stage training scheme effectively addresses SNN training difficulties, particularly for medical image segmentation tasks like hippocampus analysis.
• The hybrid approach offers a promising direction for developing efficient and accurate SNNs.
• Further research into data representation and scaling issues can further optimize SNN performance.