A Spiking Neural Network Based on Retinal Ganglion Cells for Automatic Burn Image Segmentation

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Severe burn injuries cause increased catabolism and significant fluid loss, leading to high mortality rates.
• Accurate estimation of burn wound area (as a percentage of Total Body Surface Area - TBSA%) is critical for early patient treatment and fluid management.
• Current methods for burn wound assessment suffer from observer variability, necessitating objective and accurate localization techniques.

Purpose of the Study:

• To develop an objective and accurate method for burn wound segmentation.
• To address the limitations of existing Convolutional Neural Networks (CNNs) regarding computational resources on edge hardware.
• To propose a lightweight model for efficient burn wound segmentation.

Main Methods:

• Construction of a dedicated burn image dataset.
• Development of a U-type Spiking Neural Network (SNN) model inspired by retinal ganglion cells (RGC) for segmenting burn and non-burn areas.
• Introduction of a cross-layer skip concatenation module to enhance network performance.

Main Results:

• The proposed U-type SNN model achieved a pixel accuracy of 92.89% for burn wound segmentation.
• The network requires only 16.6 Mbytes of parameters, demonstrating its lightweight nature.
• The model exhibited remarkable accuracy and suitability for edge hardware.

Conclusions:

• The developed lightweight SNN model offers an accurate and efficient solution for burn wound segmentation.
• This approach overcomes the computational constraints of traditional CNNs for edge deployment.
• The findings highlight the potential of SNNs in medical imaging for critical care applications.