CDUNeXt: efficient ossification segmentation with large kernel and dual cross gate attention

• Published in: Scientific reports 2024-12, Vol.14 (1), p.31018-14
• Main Authors: Xia, Hailiang, Wang, Chuantao, Li, Zhuoyuan, Zhang, Yuchen, Hu, Shihe, Zhai, Jiliang
• Format: Article
• Language: English
• Subjects: 639/705/117; 639/705/258; 692/308; 692/698/1671; 692/699; Algorithms; Deep Learning; Diagnosis; Dual cross gate attention; Humanities and Social Sciences; Humans; Image processing; Image Processing, Computer-Assisted - methods; Large kernel convolution; Ligamentum Flavum - diagnostic imaging; multidisciplinary; Ossification; Ossification of the ligamentum flavum; Ossification, Heterotopic - diagnostic imaging; Ossification, Heterotopic - pathology; Osteogenesis; Science; Science (multidisciplinary); Spatial discrimination learning; Spinal stenosis; Spinal Stenosis - diagnostic imaging; Spinal Stenosis - pathology; Tomography, X-Ray Computed - methods
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-024-82199-9

Ossification of the ligamentum flavum (OLF) is the main causative factor of spinal stenosis, but how to accurately and efficiently identify the ossification region is a clinical pain point and an urgent problem to be solved. Currently, we can only rely on the doctor’s subjective experience for identification, with low efficiency and large error. In this study, a deep learning method is introduced for the first time into the diagnosis of ligamentum flavum ossificans, we proposed a lightweight, automatic and efficient method for identifying ossified regions, called CDUNeXt. By designing lightweight module structures, utilizing large-kernel convolutions to extracts the long-distance dependencies of different features of the image, and adopting dual-cross-gate-attention(DCGA) to sequentially capture the channel and spatial dependencies so as to fast and accurate segmentation while maintaining fewer parameters and lower complexity. Experiments show that CDUNeXt achieves the best segmentation performance with an optimal balance of lighter weights and less computational cost compared to existing methods. This work fills the gap in the application of deep learning techniques in the diagnosis of ligamentum flavum ossificans, contributes to the realization of lightweight medical image segmentation networks and lays the foundation for subsequent research.