A Three-Stage Semi-Supervised Learning Approach to Spine Image Segmentation

Spine segmentation in computed tomography (CT) images is critical for automatic analysis, especially when focusing on varied spinal anatomy. Despite having comprehensive annotations for normal vertebrae, many datasets do not encompass labeled fracture data, posing challenges for predictive modeling. This research presents a three-stage 2.5D semi-supervised learning based on U-Net that utilizes both labeled and unlabeled datasets. The objectives are to reduce workload needed for manual annotation and create a model proficient in processing fracture data without prior specific fracture dataset with labeling. Due to the similarity between the vertebrae, precise segmentation is difficult. We utilized a cascade framework, which is aligned to a structured clinical examination process of the vertebral segments in order to achieve more precise delineation. In view of the voluminous data in 3D CT images and GPU performance constraints, this study strategically employs 2D network training, further supplemented by 2.5D network input, to optimize model performance. Preliminary findings suggest that this approach significantly improves the model's ability to segment spine regions, especially in environments with limited equipment capabilities. Further evaluation is required to understand its full potential in various scenarios, including impact on detection of fractures.