Abstract
Image registration is a fundamental task in medical imaging, yet traditional deformable algorithms often struggle with local minima, varying fields of view, and rigid parameter constraints. We introduce FLOOR (Fusing Locally Optimal Registrations), a novel framework that overcomes these limitations by generating a diverse set of candidate registrations and fusing the best local results into a single, optimized deformation map. By integrating traditional iterative methods with manifold learning, FLOOR significantly improves the accuracy of complex cardiac MRI alignments, particularly in challenging anatomical regions like the myocardium.
Introduction
Standard deformable registration algorithms, such as the popular "Demons" algorithm, face three primary challenges:
- Local Minima: Algorithms frequently converge on mathematically plausible but physically incorrect alignments.
- Missing Correspondences: Discrepancies in the field of view between scans can lead to artifacts as the algorithm attempts to force matches where none exist.
- Parameter Sensitivity: Global smoothing parameters ($\sigma$) fail to account for the varying flexibility requirements of different tissues (e.g., rigid bone vs. contractile muscle).
FLOOR addresses these issues by shifting the focus from finding a single global solution to identifying and fusing the best local solutions across multiple registration attempts.
Methods
The FLOOR framework operates through three distinct phases:
Candidate Generation
Rather than a single registration pass, FLOOR executes multiple registrations with varying smoothing parameters ($\sigma$). Crucially, the algorithm is applied to different "parcels"—overlapping sub-regions of the image. This ensures that misalignments in one area do not propagate across the entire image.
Manifold Learning (MLD)
The framework incorporates Manifold Learning based Deformation maps (MLD). This technique treats a collection of images as points on a high-dimensional manifold and calculates the geodesic (shortest path) between them to suggest alternative registration candidates.
Iterative Fusion
The "secret sauce" of FLOOR is its iterative fusion process. At each voxel, the algorithm evaluates all candidates and selects the one that minimizes the local energy function. These selections are then smoothed into a coherent global deformation map. This process repeats until convergence, creating a virtuous cycle where improved registrations lead to a more accurate manifold.
Experimental Evaluation
The study utilized a cardiac MRI dataset consisting of 25 subjects (15 healthy volunteers and 10 patients with Tetralogy of Fallot). The authors performed 600 pairwise registrations. Accuracy was measured using Dice scores based on expert manual segmentations of the Right Ventricle (RV), Left Ventricle Myocardium (MY), and Left Ventricle blood pool (LV).
Results
FLOOR demonstrated superior performance compared to standard Demons and global manifold learning (GMLD) baselines.
- Quantitative Accuracy: FLOOR achieved an average Dice score of 79.2%, significantly higher than standard Demons (73.6%) and GMLD (72.9%).
- Myocardium Performance: The most notable improvement occurred in the thin and difficult-to-register myocardium, where FLOOR reached a 68.1% Dice score, compared to 59.7% for Demons.
- Qualitative Robustness: Visual analysis confirmed that FLOOR effectively eliminated warping artifacts common in baseline methods when handling varying fields of view.
Discussion and Limitations
The core innovation of FLOOR is the realization that a globally optimal registration is effectively a collection of locally optimal ones. By allowing spatial variation in smoothing, the model maintains rigidity where necessary while remaining flexible in dynamic regions. Furthermore, parcellation confines the impact of missing correspondences, preventing global distortion.
However, the method is not without limitations. The computational cost of generating multiple candidates and training manifold models is high, potentially limiting real-time clinical application. Additionally, the framework's success depends on the quality of the initial candidate set. While promising for cardiac MRI, further research is needed to validate its generalizability to other anatomical regions such as the brain or lungs.
Summary updated