Variability in commercially available deformable image registration: A multi-institution analysis using virtual head...

Area of Science:

• Medical imaging
• Radiotherapy
• Image analysis

Background:

• Deformable image registration (DIR) is crucial for adaptive radiotherapy.
• Evaluating the performance of DIR software across institutions is essential for clinical implementation.

Purpose of the Study:

• To assess the performance of three common DIR software packages: Velocity, MIM, and Eclipse.
• To analyze the variability in DIR accuracy across different institutions and algorithms.

Main Methods:

• Utilized ten virtual phantoms from head-and-neck cancer patient CT scans.
• Collected DIR results from 35 institutions using Velocity, MIM, or Eclipse.
• Calculated target registration error (TRE) by comparing submitted deformation vector fields (DVFs) to ground-truth DVFs.

Main Results:

• MIM achieved the lowest mean TRE (1.10 ± 0.29 mm), significantly outperforming Velocity (2.04 ± 0.35 mm) and Eclipse (2.35 ± 0.15 mm).
• Significant institutional differences were observed for MIM and Velocity, potentially due to parameter variations and software versions.
• Maximum errors exceeding 2 cm were noted, highlighting the need for caution in DIR application.

Conclusions:

• All evaluated DIR software platforms showed average TREs below 3 mm.
• MIM demonstrated statistically superior performance, but all algorithms met a benchmark of being better than the largest voxel dimension.
• Algorithm choice may impact DIR accuracy more than user technique; quality assurance after software upgrades is recommended.