RegUn: Registration Uncertainty: Quantitative Error Prediction of Medical Image Registration using Regression Forests

Introduction

Predicting registration error can be useful for evaluation of registration procedures, which is important for the adoption of registration techniques in the clinic. In addition, quantitative error prediction can be helpful in improving the registration quality. This work proposes a new automatic method to predict the registration error in a quantitative manner, and is applied to chest CT scans. A random regression forest is utilized to predict the registration error locally. The forest is built with features related to the transformation model and features related to the dissimilarity after registration.

Figure 1: An example of RegUn.

Citation

[1] Sokooti, H., Saygili, G., Glocker, B., Lelieveldt, B.P. and Staring, M., 2019. Quantitative Error Prediction of Medical Image Registration using Regression Forests. Medical image analysis. arXiv

[2] Sokooti, H., Saygili, G., Glocker, B., Lelieveldt, B. P., & Staring, M. (2016, October). Accuracy estimation for medical image registration using regression forests. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 107-115). Springer, Cham.

1. Dependencies

• numpy : General purpose array-processing package.
• SciPy : A Python-based ecosystem of open-source software for mathematics, science, and engineering.
• SimpleITK : Simplified interface to the Insight Toolkit for image registration and segmentation.
• TBB: Lets you easily write parallel C++ programs that take full advantage of multicore performance.

2. Running RegUn

Check uncertainty.py

2.1 Perform Registration:

In order to perform registration, runnig the script do_elastix_registration.py or do_ANTs_registration.py is easier. Later the script uncertainty.py can read the registration results.

An example of a registration paramater for elastix and ANTs package are available at:

Elastix/DIR-Lab_COPD/elastix1/parameter/ and

Elastix/DIR-Lab_COPD/ANTs1/parameter/.

2.1.1 Perform Registration using a cluster:

The software is capble to use an Open Grid Scheduler cluster. This can be done by modifying the parameter where_to_run:

where_to_run = 'sharkCluster'

variable	Task	Number of Registration
setting['cluster_phase'] = 0	affine registration	1
setting['cluster_phase'] = 1	initial perturbation to calculate stdT	21
setting['cluster_phase'] = 2	final perturbation to calculate stdTL	20

By assigning setting['cluster_task_dependency'] = True, the software automatically waits for the earlier phases to be completed.

2.2 Reading Images

All of the addressess (images, results, etc) can be modified in Functions/Python/setting_utils.py.

2.3 Pooling

In order to calculate max-pooling, average-pooling and normalized mutual information, the binary versions are available at: Functions/Python/EXE/.