Poster Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

INTRODUCTION

Spinal tumors are a complex spinal disorder, and surgical treatment is the primary approach for managing them. Spinal tumor surgery is characterized by its high complexity, extended duration, high risks, and a significant potential for complications. Therefore, accurate preoperative planning plays a vital role in ensuring surgical safety, reducing operation time, and ensuring the success of the procedure. Currently, the preoperative planning for spinal tumor surgery is mainly carried out manually by surgeons, requiring them to manually identify and segment the tumor region. This planning method is labor-intensive, time-consuming, and greatly hinders the efficiency of clinical treatment.

Due to the specialized nature, high complexity, and time-consuming nature of annotating tumor images, medical image segmentation techniques based on fully supervised learning are challenging to apply to tumor image segmentation tasks. A semisupervised learning framework efficiently trains artificial intelligence models using partially labeled data, achieving expert-level accuracy in tumor lesion segmentation and significantly reducing model training costs. Therefore, this paper proposes a semisupervised learning-based method for spinal tumor segmentation, validated using spinal tumor CT images collected in clinical practice. This approach significantly enhances clinical efficiency, addresses the issue of lengthy traditional surgical planning, and indirectly alleviates the burden on clinical practitioners

METHODS

This paper introduces a semisupervised spinal tumor segmentation framework based on three-dimensional topological constraints and uncertainty analysis, validated on spinal tumor CT images collected in clinical settings. The framework decodes spatial topological relationships between different anatomical structures in the images and employs uncertainty analysis methods based on information entropy theory to enhance the accuracy of tumor boundary identification.

This study conducted clinical validation using a total of 50 spinal tumor CT images provided by Peking University Third Hospital. Experienced radiologists manually annotated the tumor regions in these 50 spinal tumor CT images. Out of these, 40 images (20 with labels and 20 without labels) were used for model training, 5 images were used for validation during the model training process, and the final 5 images were used to test the model's performance. The study employed DICE coefficient, accuracy, sensitivity, and specificity to measure the accuracy of the segmentation results. Additionally, the paper visualized the model's segmentation results, providing a visual comparison with the doctors' manual annotations. The study also employed the Wilcoxon signed-rank test to demonstrate that the model proposed in this paper has statistically significant advantages over existing methods.

RESULTS

On the test set, the model achieved DICE, accuracy, sensitivity, and specificity scores of 0.833, 0.981, 0.859, and 0.982, respectively. As shown in the figure below, the visual segmentation results of the proposed method exhibit no significant differences from the manual annotations by expert physicians (the red curve represents the model's segmented boundary, while the blue curve represents the manually annotated tumor boundary).

DISCUSSION

This paper established a semisupervised spinal tumor segmentation framework and validated it on spinal tumor CT images from clinical practice. Experimental results demonstrate that, compared to existing methods, this framework achieves spinal tumor segmentation results closer to those of manual annotations by human doctors.