Intelligent recognition and segmentation of anatomical structures in spinal endoscopy: a deep learning approach with 1000 annotated images.

backgroundAccurate intraoperative identification of anatomical structures is critical for ensuring the safety and efficacy of spinal endoscopic surgery. This task, however, is frequently hampered by a limited visual field, image blurring, and complex anatomical relationships inherent to the endoscopic environment. Existing intelligent systems lack robustness in handling multi-structure segmentation under such challenging conditions.

objectiveTo develop and validate a deep learning-based system for the automated, simultaneous segmentation of multiple key anatomical structures in spinal endoscopic images, aiming to provide a reliable foundation for computer-aided surgical navigation.

methodsWe constructed a large-scale, expert-annotated dataset of 1000 spinal endoscopic images. Building upon the U-Net architecture, we developed an enhanced model by innovatively integrating a Convolutional Block Attention Module (CBAM) and an Atrous Spatial Pyramid Pooling (ASPP) module to address challenges of low contrast and multi-scale anatomy. Model stability was assessed via five-fold cross-validation, and final performance was evaluated on an independent test set of 100 images. Ablation studies and comparisons against benchmark algorithms (U-Net, DeepLabV3+, Attention U-Net, TransUNet) were conducted. Statistical significance was assessed with paired t-tests.

resultsThe proposed system demonstrated high and consistent segmentation performance for the ligamentum flavum (Dice: 0.882 ± 0.015; mIoU: 0.850 ± 0.012), nerve roots (mIoU: 0.800 ± 0.018), and intervertebral disc (mIoU: 0.723 ± 0.021). Ablation studies confirmed the critical contributions of the CBAM and ASPP modules. Our model significantly outperformed baseline and state‑of‑the‑art architectures across major structures (P < 0.01, Bonferroni‑corrected), achieving a 17.0% relative improvement in mIoU for the ligamentum flavum compared to DeepLabV3+. Sensitivity analyses confirmed model robustness, and subgroup analyses revealed consistent performance across major spinal regions.

conclusionThis study presents a promising segmentation system for spinal endoscopic anatomy that achieves high accuracy for critical soft tissues such as the ligamentum flavum and nerve roots. The integration of attention mechanisms and multi-scale feature extraction proves to be an effective strategy. However, the current accuracy for posterior longitudinal ligament and bone remains insufficient for clinical use, and further refinement is required before real-time navigation can be considered. Future work will focus on enhancing the segmentation of challenging structures with low contrast or fragmented appearances.