Accuracy Estimation for Medical Image Registration Using Regression Forests

Abstract

Abstract. This paper reports a new automatic algorithm to estimate the misregistration in a quantitative manner. A random regression forest is constructed, predicting the local registration error. The forest is built using local and modality independent features related to the registration precision, the transformation model and intensity-based similarity after registration. The forest is trained and tested using manually annotated corresponding points between pairs of chest CT scans. The results show that the mean absolute error of regression is 0.72 ± 0.96 mm and the accuracy of classification in three classes (correct, poor and wrong registration) is 93.4%, comparing favorably to a competing method. In conclusion, a method was proposed that for the first time shows the feasibility of automatic registration assessment by means of regression, and promising results were obtained.

A supervised learning approach for detecting erroneous samples in embeddings

“…After the feature extraction, feature pooling is applied akin to [23,25,26]. Feature pooling enlarges feature space by extracting new features from the existing ones using the mean filter.…”

“…In this work, we aim to detect such errors in an embedding to inform the observer. Recently, detecting errors in medical image registration has been the aim of several studies [23][24][25][26]. Two of those studies use matching to extract features from two scans [24,25] and use this information in a random forest (RF) regressor to predict the errors [25].…”

Quantitative error prediction of medical image registration using regression forests

“…We set m to the square root of the total number of features in that experiment, which performed slightly better than m = (number of features)/3 (Liaw et al, 2002). The total number of registrations P is chosen as 20 to ensure that the estimation of std T does not change considerably when increasing the number of registrations (Sokooti et al, 2016).…”

“…The proposed method is applied and evaluated on chest CT scans. This work is an extension of Sokooti et al (2016) with updated methodology and substantially extended evaluation.…”

Deformable image registration using convolutional neural networks

“…4 Muenzing et al and Sokooti et al developed methods that can learn to estimate registration quality metrics for non-linear registration based on classification and regression of the registration error respectively. 5,6 Eppenhof and Pluim developed a convolutional neural networks approach for regression of registration errors. 7 This method was trained on artificial transformations applied to a small set of training images.…”