EFONet: A Deep Learning-Based Method for Lumbar Disc Segmentation in Magnetic Resonance Images

Lumbar disc degenerative disease is a prevalent lumbar condition that significantly impacts people's lives and health. The current methods for segmenting the disc region in lumbar MRI images face challenges due to the multi-tissue mixing and high similarity shown in sagittal and axial view images. These methods often produce unclear segmentation boundaries, fail to extract sufficient local information, and struggle to capture detailed semantic information like edges and textures effectively. Therefore, we propose an edge feature optimization network (EFONet) for segmenting lumbar disc based on MRI images. First, we propose an omni-dimensional selective kernel convolution (OSKC) to extract local details from multiple dimensions, addressing the issue of insufficient local information extraction in Transformer Block segmentation of the lumbar disc region. Next, we propose a dual-channel edge feature enhancement (DCEE) to enhance edge direction and position information of the lumbar disc region using edge operators. Finally, our method achieves fine-grained segmentation of lumbar disc edges in both sagittal and axial view images. The accuracy and visualization of lumbar disc segmentation were improved. Experiments demonstrate that our method outperforms all comparison networks.