[Paper Review] CTSpine1K: A Large-Scale Dataset for Spinal Vertebrae Segmentation in Computed Tomography

Yang Deng, Ce Wang+13 more|arXiv (Cornell University)|May 31, 2021

Medical Imaging and AnalysisEngineering23 references34 citations

TL;DR

CTSpine1K provides the largest publicly available annotated spine CT dataset (1005 volumes, 11k+ vertebrae) for vertebra segmentation, establishing a new benchmark and baseline using nnUnet.

ABSTRACT

Spine-related diseases have high morbidity and cause a huge burden of social cost. Spine imaging is an essential tool for noninvasively visualizing and assessing spinal pathology. Segmenting vertebrae in computed tomography (CT) images is the basis of quantitative medical image analysis for clinical diagnosis and surgery planning of spine diseases. Current publicly available annotated datasets on spinal vertebrae are small in size. Due to the lack of a large-scale annotated spine image dataset, the mainstream deep learning-based segmentation methods, which are data-driven, are heavily restricted. In this paper, we introduce a large-scale spine CT dataset, called CTSpine1K, curated from multiple sources for vertebra segmentation, which contains 1,005 CT volumes with over 11,100 labeled vertebrae belonging to different spinal conditions. Based on this dataset, we conduct several spinal vertebrae segmentation experiments to set the first benchmark. We believe that this large-scale dataset will facilitate further research in many spine-related image analysis tasks, including but not limited to vertebrae segmentation, labeling, 3D spine reconstruction from biplanar radiographs, image super-resolution, and enhancement.

Motivation & Objective

• Create a large-scale, multi-source spine CT dataset with high-quality vertebra segmentation annotations.
• Provide a unified annotation pipeline to ensure labeling consistency across sources.
• Establish baseline vertebrae segmentation performance using a strong, self-configuring model (nnUnet).
• Analyze cross-domain generalization between CTSpine1K and VerSe datasets to highlight domain gaps.

Proposed method

• Aggregate 1,005 spine CT volumes from four open sources to maximize appearance variation.
• Design a unified annotation workflow with nnUnet pre-labeling, junior/senior/clinical review, and double-checking to ensure high-quality ground truth.
• Annotate vertebrae from C1 to L6 with 25 labels and save masks in NIfTI format.
• Use 3D full-resolution nnUnet to establish a benchmarking segmentation model and report per-vertebra Dice and 95th percentile Hausdorff Distance (HD95).
• Split data into training (610), test_public (197), and test_private (198); evaluate on Test_VerSe for cross-dataset comparison.

Experimental results

Research questions

• RQ1 Can a large-scale, multi-source spine CT dataset improve vertebra segmentation by providing diverse anatomy and imaging conditions?
• RQ2 How does a strong baseline model (nnUnet) perform on CTSpine1K across different splits and on VerSe datasets?
• RQ3 What is the domain gap between CTSpine1K and public VerSe datasets in vertebra segmentation?

Key findings

• CTSpine1K contains 1,005 CT volumes with over 11,100 labeled vertebrae from diverse manufacturers and conditions.
• Mean Vertebrae DSC across Test_public, Test_private, and Test_VerSe is 0.869, 0.840, and 0.766, respectively; mean Spine DSC is 0.985, 0.984, and 0.929 respectively.
• L6 vertebrae are particularly challenging (DSC = 0 and HD = 66.21 mm for Test_public; 80.33 mm for Test_private; 45.16 mm for Test_VerSe), indicating disease and anatomical variations impact segmentation.
• The model trained on CTSpine1K achieves good performance on its own data but substantially worse performance on VerSe data, suggesting a notable domain gap.
• The COLONOG sub-dataset (within CTSpine1K) differs from VerSe COLONOG, explaining part of the domain discrepancy and highlighting the value of CTSpine1K as a complement to VerSe.
• Annotation quality and pipeline are validated by alignment with VerSe baselines and expert reviews; the dataset is intended to support broader spine image analysis tasks.