Research on a Fast Segmentation Algorithm for Brain CT Images Based on TransFuse and LSK Modules

Segmenting brain CT images is critical for medical image analysis. Traditional methods often lack accuracy in complex images, especially in areas with blurred edges or significant noise. Current segmentation algorithms also struggle with perceiving complex structures, high computational burdens, gradient disappearance, and inadequate feature representation. This study aims to tackle these challenges by optimizing brain CT image segmentation. The approach involves adjusting the structure and parameters of TransFuse to enhance feature representation, integrating LSK attention into the enhanced TransFuse for improved perception of complex structures, adding residual connections to streamline the Transformer, reducing parameters and computations, and utilizing the Smooth Maximum Unit to enhance nonlinear expression and prevent gradient disappearance. Experimental results demonstrate that the proposed algorithm achieves a segmentation accuracy of 0.7765 for brain CT images, with 101.08 MB parameters (optimal compression) and a processing speed of 51.11 frames per second, surpassing other models. The combination of these techniques enhances the accuracy of brain CT image segmentation and ensures operational efficiency of the model, providing a reliable technical solution for clinical or research-related tasks in brain CT image segmentation.