Open-Source Platform for Non-Reference-Based Quality Assessment of Image Registration
Abstract
Purpose
Deformable image registration (DIR) is integral to imaging alignment in radiation therapy, yet quality assurance (QA) on DIR performance remains challenging when reference contours are unavailable. In such situations, clinical decisions on acceptance of DIR results rely on subjective visual assessment. To address this gap, we developed an open-source, Medical Image Registration QA Toolkit (MIRQAT) integrating reference-free QA metrics to support clinical evaluation of DIR performance.
Methods
MIRQAT accepts 3D medical images with deformation vector fields (DVFs) from any DIR software. MIRQAT computes reference-free metrics, including intensity agreement measures like normalized cross-correlation (NCC), normalized mutual information (NMI), and structural similarity index (SSIM) computed on coarse grids, subdividing the images into fixed-size patches to assess local similarity with larger values indicating higher similarity. Deformation regularity metrics include Jacobian determinant (Det) the volume scaling factor (Det=1 indicating no volume change), harmonic energy (HE) which is ideally minimal, and divergence (Div)/curl quantifying rate of volume expansion and magnitude of localrotation. Metrics are interactively visualized via GUI, including overlay displays, DVF vector plots, and region‑of‑interest masking.
Results
MIRQAT was applied to 10 retrospective prostate CT–MRI DIR cases, generating QA metrics in 30.1 minutes (3.01±1.0 minutes/case). Across cases, mean Det, HE, divergence and curl magnitude were 1.00, 0.02, −0.01, and 0.20 respectively, indicating high DVF regularity, global volume preservation, and small global rotation. The mean per-case maximal patch NMI, NCC and SSIM (mean±standard deviation) were 1.10±0.02, 0.58±0.19, and 0.41±0.05 respectively, indicating poor-to-modest statistical dependence and structural similarity with patch-grids localizing similarity/dissimilarity.
Conclusion
MIRQAT integrates quantitative metrics to support efficient, routine QA of DIR workflows. The results on prostate CT-MRI DIR suggest that while transformations were generally smooth and volume preserving, they did not contain sufficient deformations. This framework complements visual inspection of DIR results, however, further evaluation with larger data is warranted.