Abstract
Constructing a probabilistic medical atlas requires selecting a reference "target" image to which all other images in a population are registered. Selecting an outlier as the target introduces geometric bias, complicating subsequent registrations. This paper presents a novel method for identifying the least biased target image by utilizing pair-wise registration, bending energy as a distance metric, and Multidimensional Scaling (MDS) to map the population's geometric distribution.
Introduction
In medical imaging, a probabilistic atlas serves as a gold-standard reference for tasks such as segmentation and registration. The construction process typically involves mapping a set of training images into a common spatial frame. A significant challenge in this process is the selection of the "anchor" or target image. If an unrepresentative or extreme case is chosen, the resulting atlas becomes biased, forcing other images to undergo unnecessary geometric transformations to reach the atlas space. Traditional iterative averaging methods to mitigate this are often computationally expensive and sensitive to noise.
Methodology
The proposed pipeline identifies the least biased target through a six-step statistical approach:
- Pair-Wise Registration: Every image in the dataset is registered to every other image ($N(N-1)/2$ total registrations) using Mutual Information (MI) and Thin-Plate Splines (TPS).
- Bending Energy Calculation: For each registration, the "bending energy" is calculated. This metric quantifies the amount of warping required to align two images, serving as a proxy for their geometric distance.
- Distance Matrix Construction: These energies are compiled into a symmetric matrix $D$, representing the geometric relationships between all pairs.
- Multidimensional Scaling (MDS): MDS projects the abstract distances from the matrix into a low-dimensional Euclidean space. In this "shape space," geometrically similar images cluster together.
- Mean Geometry Identification: The "mean geometry" of the population is determined by calculating the average coordinate of all points in the MDS space.
- Target Selection: The image located closest to this average coordinate is selected as the least biased target.
Experimental Results
The method was validated using a synthetic dataset derived from BrainWeb MRI simulations. A single undeformed slice was subjected to 50 random B-spline deformations to create a controlled population. Because the original undeformed state was known, a "ground truth" for the most central image (Image 8) was established.
The MDS method identified Image 15 as the optimal target. While the absolute ground truth was Image 8, the difference in bending energy between the two was negligible (0.0106). Furthermore, the rank-order correlation between the MDS results and the ground truth across all 50 images yielded a Root Mean Squared (RMS) error of only 0.3970, demonstrating that the MDS map accurately reflects the underlying geometric distribution.
Discussion and Conclusion
This geometric approach offers several advantages over intensity-based selection. By focusing on bending energy, the method remains robust against image noise that might otherwise skew intensity-based similarity metrics. While the $N^2$ registration requirement presents a computational challenge for very large datasets, the method is often more efficient than iterative atlas construction techniques. Future work will focus on scaling the pipeline for 3D volumes and optimizing the registration speed to handle larger clinical cohorts.
Summary updated