Efficient spine segmentation network based on multi‐scale feature extraction and multi‐dimensional spatial attention

• DOI: 10.1002/ima.23046
• Publication Date: 2024-02-27T04:15:28Z
• Authors: Guohao Xu; Chuantao Wang; Zhuoyuan Li; Jiliang Zhai; Saishuo Wang
• Keywords: Convolution (computer science); Feature (linguistics)

Abstract In spine imaging, efficient automatic segmentation is crucial for clinical decision‐making, yet current models increase accuracy at the expense of elevated parameter counts and computational complexity, complicating integration with contemporary medical devices. Addressing identified challenges, this research introduces LE‐NeXt, a framework utilizing multi‐dimensional spatial attention multi‐scale feature extraction, optimizing architecture via convolution MLP. It integrates lightweight convolutions mechanisms within an encoder‐decoder model, enhancing stage‐specific extraction while ensuring efficiency. Experimental analyses on VerSe SpineWeb datasets demonstrate that LE‐NeXt outperforms U‐NeXt, IoU from 87.7 to 89.8 VerSe, exceeds performance established networks such as U‐Net its variants. Significantly, SpineWeb, not only surpasses Trans in but also achieves considerable reduction both count complexity. These results emphasize LE‐NeXt's effectiveness improving precision efficiently, optimally balancing efficiency accuracy.