VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Radiology

Background:

• Automated vertebral labelling and segmentation are crucial for clinical decision support systems.
• Challenges include anatomical variations, diverse acquisition protocols, and limited public data.
• The Large Scale Vertebrae Segmentation Challenge (VerSe) was established to address these limitations.

Purpose of the Study:

• To benchmark algorithms for automated vertebral labelling and segmentation.
• To investigate algorithm performance variations across different levels (vertebra, scan, field of view).
• To assess algorithm generalizability to domain shifts in CT scan data.

Main Methods:

• Development and release of two large datasets (374 CT scans, 4505 annotated vertebrae).
• Benchmarking of 25 algorithms on these datasets.
• Evaluation of algorithm performance on intra- and inter-challenge datasets to assess generalizability.

Main Results:

• Performance analysis revealed that correctly identifying vertebrae with rare anatomical variations is key for algorithm success.
• Significant variations in performance were observed at vertebra, scan, and field-of-view levels.
• Top-performing algorithms showed varying degrees of generalizability to unseen data.

Conclusions:

• The VerSe challenge provided valuable insights into the state-of-the-art in automated spine image processing.
• Algorithm robustness in handling anatomical variations is critical for clinical applicability.
• Publicly available datasets and challenge results facilitate further research and development in the field.