Preventing Signal Degradation During Elastic Matching of Noisy DCE-MR Eye Images

Kishore Mosaliganti, Guang Jia, Johannes T. Heverhagen, Raghu Machiraju, Joel H. Saltz, Michael V. Knopp

Abstract

Dynamic Contrast Enhanced MRI (DCE-MRI) of the eye is a powerful tool for characterizing tumor biology through pharmacokinetic modeling. However, patient motion during the multi-minute acquisition process necessitates image registration, which can inadvertently degrade signal integrity through interpolation-induced blurring. This paper introduces an enhanced 3D elastic registration framework that combines Thirion’s Demons algorithm with anisotropic smoothing, high-pass filtering, and histogram matching. Results from 42 human eye datasets demonstrate that this approach successfully corrects motion while preserving the high-frequency signal components required for accurate physiological parameter estimation.

Introduction

In DCE-MRI, the temporal evolution of contrast agent concentration—the time-intensity plot—serves as a biological fingerprint for tumors, enabling the calculation of vascular density and permeability. A significant challenge in ocular imaging is patient motion, including breathing and eye rotation, which shifts the tumor across the image sequence.

While registration is required to align these images, the mathematical interpolation involved typically acts as a low-pass filter. This smoothing effect is detrimental in DCE-MRI as it artificially reduces peak signal intensities, leading to erroneous pharmacokinetic modeling. This study proposes a framework to mitigate such signal degradation during the elastic matching of noisy eye images, ensuring that motion correction does not come at the expense of data integrity.

Methodology

The core of the framework is built around Thirion’s Demons algorithm, a 3D elastic matching procedure that treats registration as a diffusion process. To counteract the inherent signal loss associated with this method, the authors implemented a multi-stage pipeline:

  1. Gradient Anisotropic Diffusion: A pre-processing step to reduce high-frequency "shot noise" while maintaining sharp tissue boundaries.
  2. High-Pass Filtering: Applied in the frequency domain before registration to "pre-sharpen" features, compensating for the expected blurring during the warping process.
  3. Demons Registration: The primary elastic matching step, suited for the "squishy" deformations of the eye.
  4. Histogram Matching: A post-processing step that maps the intensities of the registered image back to the original distribution, ensuring pixel values remain physically meaningful for modeling.
Figure 1: Flowchart of the proposed registration framework illustrating the sequence of pre-processing, registration, and post-processing steps.

The framework was validated using 42 human eye datasets acquired on a 1.5-T MR system. To establish a ground truth, motion-free datasets (e.g., EYE-OS-P05) were identified and subjected to synthesized elastic deformations, allowing for a direct comparison between registration results and known stationary data.

Results and Discussion

The enhanced framework significantly improved image alignment, with the average Pearson’s correlation between volumes increasing from 0.45 in unregistered data to 0.7 post-registration. Qualitatively, the method preserved delicate structures, such as the upper rim of the eye and tumor boundaries, which are often blurred by standard registration techniques.

Crucially, the preservation of signal integrity led to better pharmacokinetic model fits ($R^2$). The automated registration performed comparably to manual region-of-interest (ROI) placement by trained pathologists, while offering superior consistency across the temporal sequence. By addressing the "double-edged sword" of registration—where alignment often comes at the cost of signal smoothing—this framework provides a robust solution for ocular DCE-MRI.

Figure 6: A representative time-intensity curve (S/Spre vs. time) with a pharmacokinetic model fit, demonstrating the framework's ability to produce high-quality data for analysis.

Conclusion

This study demonstrates that integrating pre-compensation and post-processing intensity restoration into the registration pipeline can effectively prevent signal degradation in noisy DCE-MRI datasets. While the 3D elastic registration is computationally intensive, the gains in biological accuracy for tumor modeling are significant. Future work may explore the application of this intensity-preserving registration to other non-stationary organs, such as the kidneys or lungs, and investigate the impact of local versus global histogram matching on overall image quality.

Summary updated

Abstract

Motion during the acquisition of dynamic contrast enhanced MRI can cause model-fitting errors requiring co-registration. Clinical implementations use a pharmacokinetic model to determine lesion parameters from the contrast passage. The input to the model is the time-intensity plot from a region of interest (ROI) covering the lesion extent. Motion correction meanwhile involves interpolation and smoothing operations thereby affecting the time-intensity plots. This paper explores the trade-offs in applying an elastic matching procedure on the lesion detection and proposes enhancements. The method of choice is the 3D realization of the Demon’s elastic matching procedure. We validate our enhancements using synthesized deformation of stationary datasets that also serve as ground-truth. The framework is tested on 42 human eye datasets. Hence, we show that motion correction is beneficial in improving the model-fit and yet needs enhancements to correct for the intensity reductions during parameter estimation.
BibTeX
@inproceedings{mosaliganti2006preventing,
  author = {Kishore Mosaliganti and Guang Jia and Johannes T. Heverhagen and Raghu Machiraju and Joel H. Saltz and Michael V. Knopp},
  title = {{Preventing Signal Degradation During Elastic Matching of Noisy DCE-MR Eye Images}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2006},
  doi = {10.1007/11866565\_102},
}
Figure