Abstract
In image-guided neurosurgery, current commercial systems usually provide only rigid registration, partly because it is harder to predict, validate and understand non-rigid registration error. Probabilistic image registration (PIR) methods provide measures of registration uncertainty, which could be a surrogate for assessing the registration error. It is intuitive and believed by many clinicians that high uncertainty indicates a large error. However, the monotonic association between uncertainty and error has not been examined in image registration literature. In this pilot study, we investigate the relation between Gaussian process (GP) uncertainty and error based on clinical data and show empirically that there is a weak-to-moderate positive monotonic correlation between point-wise GP registration uncertainty and non-rigid registration error.
Introduction: The Challenge of Brain Shift
During brain surgery, the brain physically shifts once the skull is opened—a phenomenon known as "brain shift." To help surgeons navigate, preoperative scans are "registered" (aligned) to intraoperative ultrasound (iUS). While rigid registration is the clinical standard, it cannot account for the complex, non-linear deformations of the brain.
Non-rigid (deformable) registration is more accurate but much harder to validate in real-time. Surgeons need to know if they can trust the alignment, especially near critical structures like blood vessels or tumor margins. Probabilistic Image Registration (PIR) methods offer a potential solution by generating "uncertainty maps." The intuitive assumption is that if the algorithm is "uncertain" about a region, the error there is likely high. However, this monotonic association—where error increases as uncertainty increases—had never been rigorously proven in a neurosurgical context.
Methodology: Gaussian Process Registration
The authors focused on Gaussian Process (GP) registration, a promising PIR method that treats displacement as a stochastic process.
The study employed several statistical approaches to test the uncertainty-error link:
- Spearman’s Rank Correlation ($\rho_s$): Used to measure monotonicity—simply whether the error tends to increase when uncertainty increases, regardless of the specific "shape" of the relationship.
- Point-wise Posterior Predictive Checking: The predicted displacement of specific landmarks was compared against known ground-truth positions to calculate the actual error ($\epsilon$) and compared it to the GP-derived uncertainty ($u$).
- Patch-wise Correlation: Since surgeons often look at regions rather than single points, the authors tested the correlation over image patches using intensity-based metrics like Histogram Intersection (HI) and Sum of Squared Differences (SSD).
Datasets
The study utilized two clinical datasets containing a total of 23 pairs of preoperative and intraoperative ultrasound scans:
- RESECT: A public benchmark dataset for image-guided neurosurgery featuring manually annotated landmarks.
- MIBS: A proprietary dataset from a local hospital using automatically detected landmarks.
Key Results and Insights
The results suggest a significant gap between clinical intuition and statistical reality.
- Point-wise Results: The correlation was positive but underwhelming. For the RESECT dataset (manual landmarks), the mean $\rho_s$ was 0.2899 (weak-to-moderate). For the MIBS dataset (automatic landmarks), it was slightly higher at 0.4014.
- Patch-wise Results: The correlation scores were consistently low across the board. Using intensity metrics like HI or SSD to estimate patch error did not yield a reliable link to the GP uncertainty maps.
- The "Distance" Trap: In GP registration, uncertainty is largely a function of how far a voxel is from the nearest known landmark. If a point is far from a landmark but happens to be in a region with very little actual deformation, the algorithm will report high uncertainty despite the error being low. This "false positive" uncertainty weakens the correlation.
Conclusion
The authors conclude that at the current stage, it is not feasible to use GP registration uncertainty as a reliable surrogate for registration error in clinical practice. While there is a weak-to-moderate positive monotonic correlation, it is not strong enough to be trusted blindly in the operating room. Future work should explore regional uncertainty measures and investigate whether different PIR models or hybrid uncertainty metrics might yield a stronger association.
Summary updated