A multi-stage semi-supervised learning approach to spine image segmentation

• Main Author: Pan, Ruixiang
• Other Authors: Lin Zhiping
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Spine segmentation; Unet; Semi-supervised learning; Fracture; 2.5D; Multi-stage
• Online Access: https://hdl.handle.net/10356/174682
• Institution: Nanyang Technological University

Spine segmentation in computed tomography (CT) images is critical for automatic analysis, especially when focusing on varied spinal regions. Despite having comprehensive annotations for normal vertebrae, the current dataset does not encompass fracture data, posing challenges for predictive modeling. This research introduces a multi-stage 2.5D Unet-based semi-supervised learning methodology that utilizes both labeled and unlabeled datasets. The objective is to not only diminish the manual annotation workload but also to create a model proficient in processing fracture data without specific fracture dataset labeling. Due to the similarity between the vertebrae, precise segmentation is difficult, so considering a pipeline mimicking the clinical steps in examination of vertebral segmentation benefits the understanding of the vertebral bone condition by utilizing the cascade framework. The cumulative knowledge not only ensures a solid foundation for subsequent stages, but also promotes explanainability in detailed output. Considering the voluminous data in 3D CT images and GPU performance constraints, this study strategically employs 2D network training, further supplemented by 2.5D network training, 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 the presence of fractures.