Attention-interactive horizontal-vertical graph-aware network for medical spine segmentation

• Published in: Engineering applications of artificial intelligence 2025-03, Vol.143, p.110013, Article 110013
• Main Authors: Tian, Yuyu, Lv, Yalong, Cai, Xiaoyia
• Format: Article
• Language: English
• Subjects: Global contextual semantics; Graph representation; Local spatial semantics; Medical spine segmentation; Prior knowledge
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2025.110013

The health of the spine is vital to the human body. However, the single-scale convolutional network cannot fully capture the local detail features of the spine region. Meanwhile, the attribute decision-making ability of the boundary pixels between the spine region and the surrounding tissue structure could be better. To address these problems, we developed an attention-interactive horizontal-vertical graph-aware network to explore the semantic representation of the details of the spine region in magnetic resonance images. The network captures the coarse-grained multi-scale local spatial details of the spine region by dividing the subspace structure. Attention groups are used to refine the coarse-grained spatial features and establish rich channel dependencies and interactions between different subspace structures. Learning graph representation from horizontal and vertical directions and using the dynamic aggregation function between graph-aware nodes in the neighborhood to extract fine-grained semantics highlights the difference between the spinal region and the surrounding tissue structure. It improves boundary pixels’ attribute decision-making ability. The designed triple local attention embedding module establishes a complementary relationship. It balances coarse-grained features, fine-grained semantics, and prior knowledge to compensate for the shortcomings of a single feature representing the spine region. In addition, the designed weighted loss function performs separate supervised learning on each module and adaptively imposes penalties and constraints to encourage the network to achieve optimal performance. Finally, experimental results on the MRSpineSeg and CHAOS datasets show that the proposed framework has good segmentation performance and robustness, namely, mean intersection over union is 83.78% and 94.8%, respectively.