Author: Deshan Yang, Zhendong Zhang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose:
The evaluation of deformable image registration (DIR) algorithms is crucial for improving accuracy and clinical adoption. However, reliable benchmarks, especially for inter-modality registration, remain scarce. This study presents an automated framework that detects landmark pairs precisely at corresponding vessel bifurcations in intra-patient MR-CT images, providing a robust foundation for developing benchmark datasets to quantitatively assess DIR methods for MR-CT registration.
Methods:
Five MR-CT liver image pairs were obtained from publicly available archives and processed using a semi-automated pipeline to generate precise landmark pairs. A deep learning-based model was first employed to segment liver vessels in the MR images, followed by automated detection of vessel bifurcations as landmark candidates using morphological operations. Corresponding landmarks were then placed in the CT images through deformable image registration. To ensure accuracy, manual validation followed by manual adjustments was conducted to refine landmark positions and mitigate uncertainties arising from variations in vessel segmentation due to inconsistent image quality.
Results:
The semi-automatic landmarking pipeline was evaluated on a representative case, detecting approximately 100 landmark pairs per image pair, uniformly distributed across the liver. This significantly exceeded the number of identifiable landmarks recognized by human observers. The positional accuracy and correspondence of the detected landmarks were rigorously assessed through a comprehensive manual verification process, ensuring high reliability for DIR evaluation.
Conclusion:
The proposed framework successfully identified landmark pairs in MR-CT liver image pairs, providing a robust and efficient approach for DIR validation. This workflow will be extended to a larger dataset to establish the first benchmark dataset for multi-modality DIR evaluation, facilitating advancements in inter-modality image registration research.