Quantitative error prediction of medical image registration using regression forests

Hessam Sokooti, Gorkem Saygili, Ben Glocker, Boudewijn P. F. Lelieveldt, Marius StaringMedical Image Analysis 2019 · DOI

Introduction

Image registration—the spatial alignment of two or more images into a common coordinate system—is a fundamental task in medical imaging, critical for longitudinal tumor tracking and radiotherapy planning. Despite its importance, most registration algorithms lack built-in mechanisms to quantify their own accuracy. Current quality assurance (QA) typically relies on time-consuming manual inspection by experts, which is subjective and prone to error. This paper presents an automated framework to predict local, quantitative registration errors in physical units (millimeters) using Random Regression Forests (RF), providing a robust safety check for clinical workflows.

Methodology

The proposed system utilizes a Random Regression Forest to map a diverse set of local image features to a predicted Euclidean error. The feature set is categorized into three main groups:

  1. Registration-based Features: These assess the stability of the registration process. By perturbing the initial parameters of the registration algorithm, the system measures how consistently it converges. High variance in the resulting transformations (standard deviation of the transformation, or std T) indicates potential instability and higher error.
  2. Intensity-based Features: These evaluate the alignment quality based on image textures. Metrics such as MIND (Modality Independent Neighborhood Descriptor) and Mutual Information are employed to measure local similarity, even in the presence of varying noise or contrast levels.
  3. Feature Pooling: To incorporate spatial context, features are aggregated over multi-scale neighborhoods (ranging from 2mm to 40mm) using mean and maximum operators.
Overview of the error prediction pipeline using Random Forest regression.

The RF model is trained on ground-truth landmarks to learn the complex relationships between these multi-modal features and the actual registration error.

Experimental Evaluation

The framework was validated using three lung CT databases:

  • SPREAD: 21 pairs of follow-up 3D lung CTs.
  • DIR-Lab-4DCT: 5 cases capturing respiratory motion across 5 phases.
  • DIR-Lab-COPDgene: 10 cases involving patients with severe Chronic Obstructive Pulmonary Disease.

Testing included both intra-database scenarios and inter-database generalization tests to evaluate the model's performance across different clinical conditions and registration toolboxes (e.g., elastix and ANTs).

Results and Discussion

The model demonstrated high quantitative accuracy, achieving a Mean Absolute Error (MAE) of 1.07 ± 1.86 mm on the SPREAD dataset. In cross-database testing, the MAE remained robust at 1.76 ± 2.59 mm. Beyond continuous prediction, the model effectively categorized registrations into "Correct," "Poor," and "Wrong" with an overall accuracy of 90.7%.

Visualization of predicted error maps on a lung CT scan, showing localized regions of high registration uncertainty.

Key findings highlight that the synergy between intensity-based and registration-based features is crucial for performance. Specifically, the std T feature and MIND descriptors emerged as the most potent predictors of error. By providing a continuous error map rather than a simple binary classification, this method enables "uncertainty-aware" downstream analysis, allowing clinicians to focus their attention on specific regions of high predicted error.

Conclusion and Future Work

This research provides a scalable, automated solution for the quantitative quality assessment of medical image registration. While the current model successfully predicts error magnitude, future iterations aim to predict directional error vectors to assist in automated registration refinement. Additionally, distinguishing between registration failures and genuine physiological changes (such as tumor shrinkage) remains a key area for future development, potentially through the integration of deep learning architectures like Convolutional Neural Networks.

Summary updated

BibTeX
@article{sokooti2019quantitative,
  author = {Hessam Sokooti and Gorkem Saygili and Ben Glocker and Boudewijn P. F. Lelieveldt and Marius Staring},
  title = {{Quantitative error prediction of medical image registration using regression forests}},
  journal = {Medical Image Analysis},
  year = {2019},
  doi = {10.1016/j.media.2019.05.005},
}
Figure