Floris F. Berendsen

Multiatlas-based segmentation with preregistration atlas selection

PURPOSE Automatic, atlas-based segmentation of medical images benefits from using multiple atlases, mainly in terms of robustness. However, a large disadvantage of using multiple atlases is the large computation time that is involved in registering atlas images to the target image. This paper aims to reduce the computation load of multiatlas-based segmentation by heuristically selecting atlases before registration. METHODS To be able to select atlases, pairwise registrations are performed for all atlas combinations. Based on the results of these registrations, atlases are clustered, such that each cluster contains atlas that registers well to each other. This can all be done in a preprocessing step. Then, the representatives of each cluster are registered to the target image. The quality of the result of this registration is estimated for each of the representatives and used to decide which clusters to fully register to the target image. Finally, the segmentations of the registered images are combined into a single segmentation in a label fusion procedure. RESULTS The authors perform multiatlas segmentation once with postregistration atlas selection and once with the proposed preregistration method, using a set of 182 segmented atlases of prostate cancer patients. The authors performed the full set of 182 leave-one-out experiments and in each experiment compared the result of the atlas-based segmentation procedure to the known segmentation of the atlas that was chosen as a target image. The results show that preregistration atlas selection is slightly less accurate than postregistration atlas selection, but this is not statistically significant. CONCLUSIONS Based on the results the authors conclude that the proposed method is able to reduce the number of atlases that have to be registered to the target image with 80% on average, without compromising segmentation accuracy.

Free-form image registration regularized by a statistical shape model: application to organ segmentation in cervical MR

Deformable registration is prone to errors when it involves large and complex deformations, since the procedure can easily end up in a local minimum. To reduce the number of local minima, and thus the risk of misalignment, regularization terms based on prior knowledge can be incorporated in registration. We propose a regularization term that is based on statistical knowledge of the deformations that are to be expected. A statistical model, trained on the shapes of a set of segmentations, is integrated as a penalty term in a free-form registration framework. For the evaluation of our approach, we perform inter-patient registration of MR images, which were acquired for planning of radiation therapy of cervical cancer. The manual delineations of structures such as the bladder and the clinical target volume are available. For both structures, leave-one-patient-out registration experiments were performed. The propagated atlas segmentations were compared to the manual target segmentations by Dice similarity and Hausdorff distance. Compared with registration without the use of statistical knowledge, the segmentations were significantly improved, by 0.1 in Dice similarity and by 8mm Hausdorff distance on average for both structures.

End-to-End Unsupervised Deformable Image Registration with a Convolutional Neural Network

In this work we propose a deep learning network for deformable image registration (DIRNet). The DIRNet consists of a convolutional neural network (ConvNet) regressor, a spatial transformer, and a resampler. The ConvNet analyzes a pair of fixed and moving images and outputs parameters for the spatial transformer, which generates the displacement vector field that enables the resampler to warp the moving image to the fixed image. The DIRNet is trained end-to-end by unsupervised optimization of a similarity metric between input image pairs. A trained DIRNet can be applied to perform registration on unseen image pairs in one pass, thus non-iteratively. Evaluation was performed with registration of images of handwritten digits (MNIST) and cardiac cine MR scans (Sunnybrook Cardiac Data). The results demonstrate that registration with DIRNet is as accurate as a conventional deformable image registration method with short execution times.

Fully automatic segmentation of left atrium and pulmonary veins in late gadolinium‐enhanced MRI: Towards objective atrial scar assessment

To realize objective atrial scar assessment, this study aimed to develop a fully automatic method to segment the left atrium (LA) and pulmonary veins (PV) from late gadolinium‐enhanced (LGE) magnetic resonance imaging (MRI). The extent and distribution of atrial scar, visualized by LGE‐MRI, provides important information for clinical treatment of atrial fibrillation (AF) patients.

Nonrigid Image Registration Using Multi-scale 3D Convolutional Neural Networks

In this paper we propose a method to solve nonrigid image registration through a learning approach, instead of via iterative optimization of a predefined dissimilarity metric. We design a Convolutional Neural Network (CNN) architecture that, in contrast to all other work, directly estimates the displacement vector field (DVF) from a pair of input images. The proposed RegNet is trained using a large set of artificially generated DVFs, does not explicitly define a dissimilarity metric, and integrates image content at multiple scales to equip the network with contextual information. At testing time nonrigid registration is performed in a single shot, in contrast to current iterative methods. We tested RegNet on 3D chest CT follow-up data. The results show that the accuracy of RegNet is on par with a conventional B-spline registration, for anatomy within the capture range. Training RegNet with artificially generated DVFs is therefore a promising approach for obtaining good results on real clinical data, thereby greatly simplifying the training problem. Deformable image registration can therefore be successfully casted as a learning problem.

Registration of structurally dissimilar images in MRI-based brachytherapy.

A serious challenge in image registration is the accurate alignment of two images in which a certain structure is present in only one of the two. Such topological changes are problematic for conventional non-rigid registration algorithms. We propose to incorporate in a conventional free-form registration framework a geometrical penalty term that minimizes the volume of the missing structure in one image. We demonstrate our method on cervical MR images for brachytherapy. The intrapatient registration problem involves one image in which a therapy applicator is present and one in which it is not. By including the penalty term, a substantial improvement in the surface distance to the gold standard anatomical position and the residual volume of the applicator void are obtained. Registration of neighboring structures, i.e. the rectum and the bladder is generally improved as well, albeit to a lesser degree.

SimpleElastix: A User-Friendly, Multi-lingual Library for Medical Image Registration

In this paper we present SimpleElastix, an extension of SimpleITK designed to bring the Elastix medical image registration library to a wider audience. Elastix is a modular collection of robust C++ image registration algorithms that is widely used in the literature. However, its command-line interface introduces overhead during prototyping, experimental setup, and tuning of registration algorithms. By integrating Elastix with SimpleITK, Elastix can be used as a native library in Python, Java, R, Octave, Ruby, Lua, Tcl and C# on Linux, Mac and Windows. This allows Elastix to intregrate naturally with many development environments so the user can focus more on the registration problem and less on the underlying C++ implementation. As means of demonstration, we show how to register MR images of brains and natural pictures of faces using minimal amount of code. SimpleElastix is open source, licensed under the permissive Apache License Version 2.0 and available at https://github.com/kaspermarstal/SimpleElastix.

Robust motion correction for myocardial T1 and extracellular volume mapping by principle component analysis-based groupwise image registration

Myocardial tissue characterization by MR T1 and extracellular volume (ECV) mapping has demonstrated clinical value. The modified Look–Locker inversion recovery (MOLLI) sequence is a standard mapping technique, but its quality can be negatively affected by motion.

Fully automated segmentation of left atrium and pulmonary veins in late gadolinium enhanced MRI

Background Late gadolinium enhanced (LGE) MRI enables scar assessment in the left atrial (LA) wall, providing valuable information for treatment planning and image based procedure guidance in patient with atrial fibrillation (AF). Accurate and objective segmentation of the anatomy of the LA and pulmonary veins (PV) is an important prerequisite for atrial scar assessment. However, segmentation of the LA and PV’s from LGE MRI is highly complex due to morphologic variations and limited image contrast. The study aims to develop a fully automated method for LA and PV endocardial wall segmentation from LGE MRI.

Simultaneous pairwise registration for image mosaicing of TEE data

Due to the limited field-of-view of transesophageal echocardiography (TEE) images, mosaicing is required to visualize the entire left atrium in a single image. However, the small overlap between the images and the lack of a single reference image challenges the registration. Our approach is to exploit overlap of an image with multiple other images by simultaneous pairwise registration. Three images were registered to a floating common reference using a rigid transformation. The images iteratively serve as floating reference for the other images. Averaging the resulting transformations for each image will make the simultaneous registration converge to a common reference space. It was shown on randomly transformed MR brain and TEE images that the simultaneous method achieved higher success rates than regular pairwise registration. Initial results on TEE images of the left atrium demonstrated the ability of our method to register the images to a common space.