A Parallel Convolutional Network Based on Spiking Neural Systems

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Imaging

Background:

• Deep convolutional neural networks excel at image segmentation.
• Spiking neural networks offer unique nonlinear mechanisms.
• Accurate medical image segmentation is crucial for diagnosis and treatment.

Purpose of the Study:

• To introduce a novel U-shaped convolutional neural network, SPC-Net, inspired by nonlinear spiking neural P (NSNP) systems.
• To enhance feature representation and spatial detail utilization in segmentation tasks.
• To improve multi-scale contextual information extraction and reduce information loss.

Main Methods:

• Developed an SNP-like convolutional neuron structure.
• Constructed the SPC-Net incorporating dual-convolution concatenate (DCC) and dual-convolution addition (DCA) blocks.
• Implemented a dual-scale pooling (DSP) module in the network bottleneck.
• Applied and evaluated SPC-Net on the GlaS and CRAG medical image segmentation datasets.

Main Results:

• SPC-Net achieved a 90.77% DICE coefficient and 83.76% IoU score on medical image segmentation tasks.
• The model demonstrated strong performance with an 83.93% F1 score and 86.33% ObjDice coefficient.
• Experimental results indicate superior segmentation performance compared to recent methods.

Conclusions:

• The proposed SPC-Net, utilizing an SNP-like structure and novel network blocks, achieves high accuracy in medical image segmentation.
• The integration of parallel convolutions and multi-scale pooling enhances feature representation and contextual understanding.
• SPC-Net represents a significant advancement in automated medical image analysis.