Search

Motion correction of dynamic contrast enhanced magnetic resonance images (DCE-MRI) is a challenging task, due to changes in image appearance. In this study a groupwise registration, using a principle component analysis (PCA) based metric,1 is evaluated for clinical DCE MRI of the liver. The groupwise registration transforms the images to a common space, rather than to a reference volume as conventional pairwise methods do, and computes the similarity metric on all volumes simultaneously. This groupwise registration method is compared to a pairwise approach using a mutual information metric. Clinical DCE MRI of the abdomen of eight patients were included. Per patient one lesion in the liver was manually segmented in all temporal images (N=16). The registered images were compared for accuracy, spatial and temporal smoothness after transformation, and lesion volume change. Compared to a pairwise method or no registration, groupwise registration provided better alignment. In our recently started clinical study groupwise registered clinical DCE MRI of the abdomen of nine patients were scored by three radiologists. Groupwise registration increased the assessed quality of alignment. The gain in reading time for the radiologist was estimated to vary from no difference to almost a minute. A slight increase in reader confidence was also observed. Registration had no added value for images with little motion. In conclusion, the groupwise registration of DCE MR images results in better alignment than achieved by pairwise registration, which is beneficial for clinical assessment.

Breast conserving surgery followed by radiotherapy is the standard of care for early-stage breast cancer patients. Deformable image registration (DIR) can in principle be of great value for accurate localization of the original tumor site to optimize breast irradiation after surgery. However, current state-of-the-art DIR methods are not very successful when tissue is present in one image but not in the other (i.e., in case of content mismatch). To tackle this challenge, we combined a multi-objective DIR approach with simulated tissue removal. Parameters defining the area to be removed as well as key DIR parameters (that are often tuned manually for each DIR case) are determined by a multi-objective optimization process. In multi-objective optimization, not one, but a set of solutions is found, that represent high-quality trade-offs between objectives of interest. We used three state-of-the-art multi-objective evolutionary algorithms as meta-optimizers to search for the optimal parameters, and tested our approach on four test cases of computed tomography (CT) images of breast cancer patients before and after surgery. Results show that using meta-optimization with simulated tissue removal improves the performance of DIR. This way, sets of high-quality solutions could be obtained with a mean target registration error of 2.4 mm over four test cases and an estimated excised volume that is within 20% from the measured volume of the surgical resection specimen.

Discomfort detection for infants is essential in the healthcare domain, since infants lack the ability to verbalize their pain and discomfort. In this paper, we propose a robust and generic discomfort detection for infants by exploiting a novel and efficient initialization method for facial landmark localization, using an unsupervised rollangle estimation. The roll-angle estimation is achieved by fitting a 1st-order B-spline model to facial features obtained from the scaled-normalized Laplacian of the Gaussian operator. The proposed method can be adopted both for daylight and infrared-light images and supports real-time implementation. Experimental results have shown that the proposed method improves the performance of discomfort detection by 6.0% and 4.2% for the AUC and AP using daylight images, together with 6.9% and 3.8% for infrared-light images, respectively.

In images of the corneal endothelium (CE) acquired by specular microscopy, endothelial cells are commonly only visible in a part of the image due to varying contrast, mainly caused by challenging imaging conditions as a result of a strongly curved endothelium. In order to estimate the morphometric parameters of the corneal endothelium, the analyses need to be restricted to trustworthy regions - the region of interest (ROI) - where individual cells are discernible. We developed an automatic method to find the ROI by Dense U-nets, a densely connected network of convolutional layers. We tested the method on a heterogeneous dataset of 140 images, which contains a large number of blurred, noisy, and/or out of focus images, where the selection of the ROI for automatic biomarker extraction is vital. By using edge images as input, which can be estimated after retraining the same network, Dense U-net detected the trustworthy areas with an accuracy of 98.94% and an area under the ROC curve (AUC) of 0.998, without being affected by the class imbalance (9:1 in our dataset). After applying the estimated ROI to the edge images, the mean absolute percentage error (MAPE) in the estimated endothelial parameters was 0.80% for ECD, 3.60% for CV, and 2.55% for HEX.

Deformable image registration is often a slow process when using conventional methods. To speed up deformable registration, there is growing interest in using convolutional neural networks. They are comparatively fast and can be trained to estimate full-resolution deformation fields directly from pairs of images. Because deep learning-based registration methods often require rigid or affine pre-registration of the images, they do not perform true end-to-end image registration. To address this, we propose a progressive training method for end-to-end image registration with convolutional networks. The network is first trained to find large deformations at a low resolution using a smaller part of the full architecture. The network is then gradually expanded during training by adding higher resolution layers that allow the network to learn more fine-grained deformations from higher resolution data. By starting at a lower resolution, the network is able to learn larger deformations more quickly at the start of training, making pre-registration redundant. We apply this method to pulmonary CT data, and use it to register inhalation to exhalation images. We train the network using the CREATIS pulmonary CT data set, and apply the trained network to register the DIRLAB pulmonary CT data set. By computing the target registration error at corresponding landmarks we show that the error for end-to-end registration is significantly reduced by using progressive training, while retaining sub-second registration times.

Most MRI liver segmentation methods use a structural 3D scan as input, such as a T1 or T2 weighted scan. Segmentation performance may be improved by utilizing both structural and functional information, as contained in dynamic contrast enhanced (DCE) MR series. Dynamic information can be incorporated in a segmentation method based on convolutional neural networks in a number of ways. In this study, the optimal input configuration of DCE MR images for convolutional neural networks (CNNs) is studied. The performance of three different input configurations for CNNs is studied for a liver segmentation task. The three configurations are I) one phase image of the DCE-MR series as input image; II) the separate phases of the DCE-MR as input images; and III) the separate phases of the DCE-MR as channels of one input image. The three input configurations are fed into a dilated fully convolutional network and into a small U-net. The CNNs were trained using 19 annotated DCE-MR series and tested on another 19 annotated DCE-MR series. The performance of the three input configurations for both networks is evaluated against manual annotations. The results show that both neural networks perform better when the separate phases of the DCE-MR series are used as channels of an input image in comparison to one phase as input image or the separate phases as input images. No significant difference between the performances of the two network architectures was found for the separate phases as channels of an input image., Submitted to SPIE Medical Imaging 2019

Localization of cardiac anatomical landmarks is an important step towards a more robust and accurate analysis of the heart. A fully automatic hybrid framework is proposed that detects key landmark locations in cardiac magnetic resonance (MR) images. Our method is trained and evaluated for the detection of mitral valve points on long-axis MRI and RV insert points in short-axis MRI. The framework incorporates four key modules for the localization of the landmark points. The first module crops the MR image around the heart using a convolutional neural network (CNN). The second module employs a U-Net to obtain an efficient feature representation of the cardiac image, as well as detect a preliminary location of the landmark points. In the third module, the feature representation of a cardiac image is processed with a Recurrent Neural Network (RNN). The RNN leverages either spatial or temporal dynamics from neighboring slides in time or space and obtains a second prediction for the landmark locations. In the last module the two predictions from the U-Net and RNN are combined and final locations for the landmarks are extracted. The framework is separately trained and evaluated for the localization of each landmark, it achieves a final average error of 2.87 mm for the mitral valve points and an average error of 3.64 mm for the right ventricular insert points. Our method shows that the use of a recurrent neural network for the modeling of additional temporal or spatial dependencies improves localization accuracy and achieves promising results.

Volumetric Laser Endomicroscopy (VLE) is a promising balloon-based imaging technique for detecting early neoplasia in Barretts Esophagus. Especially Computer Aided Detection (CAD) techniques show great promise compared to medical doctors, who cannot reliably find disease patterns in the noisy VLE signal. However, an essential pre-processing step for the CAD system is tissue segmentation. At present, tissue is segmented manually but is not scalable for the entire VLE scan consisting of 1,200 frames of 4,096 × 2,048 pixels. Furthermore, the current CAD methods cannot use the VLE scans to their full potential, as only a small segment of the esophagus is selected for further processing, while an automated segmentation system results in significantly more available data. This paper explores the possibility of automatically segmenting relevant tissue for VLE scans using FusionNet and a domain-specific loss function. The contribution of this work is threefold. First, we propose a tissue segmentation algorithm for VLE scans. Second, we introduce a weighted ground truth that exploits the signal-to-noise ratio characteristics of the VLE data. Third, we compare our algorithm segmentation against two additional VLE experts. The results show that our algorithm annotations are indistinguishable from the expert annotations and therefore the algorithm can be used as a preprocessing step for further classification of the tissue.

Current state-of-the-art deep learning segmentation methods have not yet made a broad entrance into the clinical setting in spite of high demand for such automatic methods. One important reason is the lack of reliability caused by models that fail unnoticed and often locally produce anatomically implausible results that medical experts would not make. This paper presents an automatic image segmentation method based on (Bayesian) dilated convolutional networks (DCNN) that generate segmentation masks and spatial uncertainty maps for the input image at hand. The method was trained and evaluated using segmentation of the left ventricle (LV) cavity, right ventricle (RV) endocardium and myocardium (Myo) at end-diastole (ED) and end-systole (ES) in 100 cardiac 2D MR scans from the MICCAI 2017 Challenge (ACDC). Combining segmentations and uncertainty maps and employing a human-in-the-loop setting, we provide evidence that image areas indicated as highly uncertain regarding the obtained segmentation almost entirely cover regions of incorrect segmentations. The fused information can be harnessed to increase segmentation performance. Our results reveal that we can obtain valuable spatial uncertainty maps with low computational effort using DCNNs., This work has been submitted to SPIE 2019 conference

Segmentation and identification of the vertebrae in CT images are important steps for automatic analysis of the spine. This paper presents an automatic method based on iterative convolutional neural networks. These utilize the inherent order of the vertebral column to simplify the detection problem, so that the network can be trained with as little as ten manual reference segmentations. Vertebrae are segmented and identified one- by-one in sequential order, using an iterative procedure. Vertebrae are first roughly localized and identified in low-resolution images that enable the analysis of context information, and afterwards reanalyzed in the original high-resolution images to obtain a fine segmentation. The method was trained and evaluated with 15 spine CT scans from the MICCAI CSI 2014 workshop challenge. These scans cover the whole thoracic and lumbar part of the spine of healthy young adults. In contrast to a non-iterative convolutional neural network, which made labeling mistakes, the proposed iterative method correctly identified all vertebrae. Our method achieved a mean Dice coefficient of 0.948 and a mean surface distance of 0.29 mm and thus outperforms the best method that participated in the original challenge.

Currently, non-rigid image registration algorithms are too computationally intensive to use in time-critical applications. Existing implementations that focus on speed typically address this by either parallelization on GPU-hardware, or by introducing methodically novel techniques into CPU-oriented algorithms. Stochastic gradient descent (SGD) optimization and variations thereof have proven to drastically reduce the computational burden for CPU-based image registration, but have not been successfully applied in GPU hardware due to its stochastic nature. This paper proposes 1) NiftyRegSGD, a SGD optimization for the GPU-based image registration tool NiftyReg, 2) random chunk sampler, a new random sampling strategy that better utilizes the memory bandwidth of GPU hardware. Experiments have been performed on 3D lung CT data of 19 patients, which compared NiftyRegSGD (with and without random chunk sampler) with CPU-based elastix Fast Adaptive SGD (FASGD) and NiftyReg. The registration runtime was 21.5s, 4.4s and 2.8s for elastix-FASGD, NiftyRegSGD without, and NiftyRegSGD with random chunk sampling, respectively, while similar accuracy was obtained. Our method is publicly available at https://github.com/SuperElastix/NiftyRegSGD.

Taking a multi-objective optimization approach to deformable image registration has recently gained attention, because such an approach removes the requirement of manually tuning the weights of all the involved objectives. Especially for problems that require large complex deformations, this is a non-trivial task. From the resulting Pareto set of solutions one can then much more insightfully select a registration outcome that is most suitable for the problem at hand. To serve as an internal optimization engine, currently used multi-objective algorithms are competent, but rather inefficient. In this paper we largely improve upon this by introducing a multi-objective real-valued adaptation of the recently introduced Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) for discrete optimization. In this work, GOMEA is tailored specifically to the problem of deformable image registration to obtain substantially improved efficiency. This improvement is achieved by exploiting a key strength of GOMEA: iteratively improving small parts of solutions, allowing to faster exploit the impact of such updates on the objectives at hand through partial evaluations. We performed experiments on three registration problems. In particular, an artificial problem containing a disappearing structure, a pair of pre- and post-operative breast CT scans, and a pair of breast MRI scans acquired in prone and supine position were considered. Results show that compared to the previously used evolutionary algorithm, GOMEA obtains a speed-up of up to a factor of ∼1600 on the tested registration problems while achieving registration outcomes of similar quality.

Optical coherence tomography (OCT) yields high-resolution, three-dimensional images of the retina. A better understanding of retinal nerve fiber bundle (RNFB) trajectories in combination with visual field data may be used for future diagnosis and monitoring of glaucoma. However, manual tracing of these bundles is a tedious task. In this work, we present an automatic technique to estimate the orientation of RNFBs from volumetric OCT scans. Our method consists of several steps, starting from automatic segmentation of the RNFL. Then, a stack of en face images around the posterior nerve fiber layer interface was extracted. The image showing the best visibility of RNFB trajectories was selected for further processing. After denoising the selected en face image, a semblance structure-oriented filter was applied to probe the strength of local linear structure in a discrete set of orientations creating an orientation space. Gaussian filtering along the orientation axis in this space is used to find the dominant orientation. Next, a confidence map was created to supplement the estimated orientation. This confidence map was used as pixel weight in normalized convolution to regularize the semblance filter response after which a new orientation estimate can be obtained. Finally, after several iterations an orientation field corresponding to the strongest local orientation was obtained. The RNFB orientations of six macular scans from three subjects were estimated. For all scans, visual inspection shows a good agreement between the estimated orientation fields and the RNFB trajectories in the en face images. Additionally, a good correlation between the orientation fields of two scans of the same subject was observed. Our method was also applied to a larger field of view around the macula. Manual tracing of the RNFB trajectories shows a good agreement with the automatically obtained streamlines obtained by fiber tracking.

In Dynamic Contrast-Enhanced MRI (DCE-MRI) of the liver, a series of images is acquired over a period of 20 minutes. Due to the patient's breathing, the liver is subject to a substantial displacement between acquisitions. Furthermore, due to its location in the abdomen, the liver also undergoes marked deformation. The large deformations combined with variation in image contrast make accurate liver registration challenging. We present a registration framework that incorporates a liver segmentation to improve the registration accuracy. The segmented liver serves as region-of-interest to our in-house developed registration method called ALOST (autocorrelation of local image structure). ALOST is a continuous optimization method that uses local phase features to overcome space-variant intensity distortions. The proposed framework can confine the solution field to the liver and allow for ALOST to obtain a more accurate solution. For the segmentation part, we use a level-set method to delineate the liver in a so-called contrast enhancement map. This map is obtained by computing the difference between the last and registered first volume from the DCE series. Subsequently, we slightly dilate the segmentation, and apply it as the mask to the other DCE-MRI volumes during registration. It is shown that the registration result becomes more accurate compared with the original ALOST approach.

Quantification of the spin-lattice relaxation time, T1, of tissues is important for characterization of tissues in clinical magnetic resonance imaging (MRI). In T1 mapping, T1 values are estimated from a set of T1-weighted MRI images. Due to the limited spatial resolution of the T1-weighted images, one voxel might consist of two tissues, causing partial volume effects (PVE). In conventional mono-exponential T1 estimation, these PVE result in systematic errors in the T1 map. To account for PVE, single-voxel bi-exponential estimators have been suggested. Unfortunately, in general, they suffer from low accuracy and precision. In this work, we propose a joint multi-voxel bi-exponential T1 estimator (JMBE) and compare its performance to a single-voxel bi-exponential T1 estimator (SBE). Results show that, in contrast to the SBE, and for clinically achievable single-voxel SNRs, the JMBE is accurate and efficient if four or more neighboring voxels are used in the joint estimation framework. This illustrates that, for clinically realistic SNRs, accurate results for quantitative biexponential T1 estimation are only achievable if information of neighboring voxels is incorporated.

Optical coherence tomography (OCT) is used to produce high-resolution three-dimensional images of the retina, which permit the investigation of retinal irregularities. In dry age-related macular degeneration (AMD), a chronic eye disease that causes central vision loss, disruptions such as drusen and changes in retinal layer thicknesses occur which could be used as biomarkers for disease monitoring and diagnosis. Due to the topology disrupting pathology, existing segmentation methods often fail. Here, we present a solution for the segmentation of retinal layers in dry AMD subjects by extending our previously presented loosely coupled level sets framework which operates on attenuation coefficients. In eyes affected by AMD, Bruch’s membrane becomes visible only below the drusen and our segmentation framework is adapted to delineate such a partially discernible interface. Furthermore, the initialization stage, which tentatively segments five interfaces, is modified to accommodate the appearance of drusen. This stage is based on Dijkstra's algorithm and combines prior knowledge on the shape of the interface, gradient and attenuation coefficient in the newly proposed cost function. This prior knowledge is incorporated by varying the weights for horizontal, diagonal and vertical edges. Finally, quantitative evaluation of the accuracy shows a good agreement between manual and automated segmentation.

Localization of anatomical regions of interest (ROIs) is a preprocessing step in many medical image analysis tasks. While trivial for humans, it is complex for automatic methods. Classic machine learning approaches require the challenge of hand crafting features to describe differences between ROIs and background. Deep convolutional neural networks (CNNs) alleviate this by automatically finding hierarchical feature representations from raw images. We employ this trait to detect anatomical ROIs in 2D image slices in order to localize them in 3D. In 100 low-dose non-contrast enhanced non-ECG synchronized screening chest CT scans, a reference standard was defined by manually delineating rectangular bounding boxes around three anatomical ROIs — heart, aortic arch, and descending aorta. Every anatomical ROI was automatically identified using a combination of three CNNs, each analyzing one orthogonal image plane. While single CNNs predicted presence or absence of a specific ROI in the given plane, the combination of their results provided a 3D bounding box around it. Classification performance of each CNN, expressed in area under the receiver operating characteristic curve, was ≥0.988. Additionally, the performance of ROI localization was evaluated. Median Dice scores for automatically determined bounding boxes around the heart, aortic arch, and descending aorta were 0.89, 0.70, and 0.85 respectively. The results demonstrate that accurate automatic 3D localization of anatomical structures by CNN-based 2D image classification is feasible.

A radiotherapy (RT) treatment can last for several weeks. In that time organ motion and shape changes introduce uncertainty in dose application. Monitoring and quantifying the change can yield a more precise irradiation margin definition and thereby reduce dose delivery to healthy tissue and adjust tumor targeting. Deformable image registration (DIR) has the potential to fulfill this task by calculating a deformation field (DF) between a planning CT and a repeated CT of the altered anatomy. Application of the DF on the original contours yields new contours that can be used for an adapted treatment plan. DIR is a challenging method and therefore needs careful user interaction. Without a proper graphical user interface (GUI) a misregistration cannot be easily detected by visual inspection and the results cannot be fine-tuned by changing registration parameters. To provide a DIR algorithm with such a GUI available for everyone, we created the extension Featurelet-Registration for the open source software p atform 3D Slicer. The registration logic is an upgrade of an in-house-developed DIR method, which is a featurelet-based piecewise rigid registration. The so called "featurelets" are equally sized rectangular subvolumes of the moving image which are rigidly registered to rectangular search regions on the fixed image. The output is a deformed image and a deformation field. Both can be visualized directly in 3D Slicer facilitating the interpretation and quantification of the results. For validation of the registration accuracy two deformable phantoms were used. The performance was benchmarked against a demons algorithm with comparable results.

Multimodal groupwise registration has been of growing interest to the image processing community due to developments in scanner technologies (e.g. multiparametric MRI, DCE-CT or PET-MR) that increased both the number of modalities and number of images under consideration. In this work a novel methodology is presented for multimodal groupwise registration that is based on Laplacian eigenmaps, a nonlinear dimensionality reduction technique. Compared to recently proposed dissimilarity metrics based on principal component analysis, the proposed metric should enable a better capture of the intensity relationships between different images in the group. The metric is constructed to be the second smallest eigenvalue from the eigenvector problem defined in Laplacian eigenmaps. The method was validated in three distinct experiments: a non-linear synthetic registration experiment, the registration of quantitative MRI data of the carotid artery, and the registration of multimodal data of the brain (RIRE). The results show increased accuracy and robustness compared to other state-of-the-art groupwise registration methodologies.

Dynamic contrast-enhanced computed tomography (DCE-CT) is an emerging radiological technique, which consists in acquiring a rapid sequence of CT images, shortly after the injection of an intravenous contrast agent. The passage of the contrast agent in a tissue results in a varying CT intensity over time, recorded in time-attenuation curves (TACs), which can be related to the contrast supplied to that tissue via the supplying artery to estimate the local perfusion and permeability characteristics. The time delay between the arrival of the contrast bolus in the feeding artery and the tissue of interest, called the bolus arrival time (BAT), needs to be determined accurately to enable reliable perfusion analysis. Its automated identification is however highly sensitive to noise. We propose an accurate and efficient method for estimating the BAT from DCE-CT images. The method relies on a piecewise linear TAC model with four segments and suitable parameter constraints for limiting the range of possible values. The model is fitted to the acquired TACs in a multiresolution fashion using an iterative optimization approach. The performance of the method was evaluated on simulated and real perfusion data of lung and rectum tumours. In both cases, the method was found to be stable, leading to average accuracies in the order of the temporal resolution of the dynamic sequence. For reasonable levels of noise, the results were found to be comparable to those obtained using a previously proposed method, employing a full search algorithm, but requiring an order of magnitude more computation time.

Purpose: To develop, test, evaluate and apply a novel tool for the white matter fiber-based analysis of T1w/T2w ratio maps quantifying myelin content. Background: The cerebral white matter in the human brain develops from a mostly non-myelinated state to a nearly fully mature white matter myelination within the first few years of life. High resolution T1w/T2w ratio maps are believed to be effective in quantitatively estimating myelin content on a voxel-wise basis. We propose the use of a fiber-tract-based analysis of such T1w/T2w ratio data, as it allows us to separate fiber bundles that a common regional analysis imprecisely groups together, and to associate effects to specific tracts rather than large, broad regions. Methods: We developed an intuitive, open source tool to facilitate such fiber-based studies of T1w/T2w ratio maps. Via its Graphical User Interface (GUI) the tool is accessible to non-technical users. The framework uses calibrated T1w/T2w ratio maps and a prior fiber atlas as an input to generate profiles of T1w/T2w values. The resulting fiber profiles are used in a statistical analysis that performs along-tract functional statistical analysis. We applied this approach to a preliminary study of early brain development in neonates. Results: We developed an open-source tool for the fiber based analysis of T1w/T2w ratio maps and tested it in a study of brain development.

Whole-body diffusion-weighted (WB-DW) MRI in combination with anatomical MRI has shown a great poten- tial in bone and soft tissue tumour detection, evaluation of lymph nodes and treatment response assessment. Because of the vast body coverage, whole-body MRI is acquired in separate stations, which are subsequently combined into a whole-body image. However, inter-station and inter-modality image misalignments can occur due to image distortions and patient motion during acquisition, which may lead to inaccurate representations of patient anatomy and hinder visual assessment. Automated and accurate whole-body image formation and alignment of the multi-modal MRI images is therefore crucial. We investigated several registration approaches for the formation or stitching of the whole-body image stations, followed by a deformable alignment of the multi- modal whole-body images. We compared a pairwise approach, where diffusion-weighted (DW) image stations were sequentially aligned to a reference station (pelvis), to a groupwise approach, where all stations were simultaneously mapped to a common reference space while minimizing the overall transformation. For each, a choice of input images and corresponding metrics was investigated. Performance was evaluated by assessing the quality of the obtained whole-body images, and by verifying the accuracy of the alignment with whole-body anatomical sequences. The groupwise registration approach provided the best compromise between the formation of WB- DW images and multi-modal alignment. The fully automated method was found to be robust, making its use in the clinic feasible.

In radiotherapy treatment planning that is only based on magnetic resonance imaging (MRI), the electron density information usually obtained from computed tomography (CT) must be derived from the MRI by synthesizing a so-called pseudo CT (pCT). This is a non-trivial task since MRI intensities are neither uniquely nor quantitatively related to electron density. Typical approaches involve either a classification or regression model requiring specialized MRI sequences to solve intensity ambiguities, or an atlas-based model necessitating multiple registrations between atlases and subject scans. In this work, we explore a machine learning approach for creating a pCT of the pelvic region from conventional MRI sequences without using atlases. We use a random forest provided with information about local texture, edges and spatial features derived from the MRI. This helps to solve intensity ambiguities. Furthermore, we use the concept of auto-context by sequentially training a number of classification forests to create and improve context features, which are finally used to train a regression forest for pCT prediction. We evaluate the pCT quality in terms of the voxel-wise error and the radiologic accuracy as measured by water-equivalent path lengths. We compare the performance of our method against two baseline pCT strategies, which either set all MRI voxels in the subject equal to the CT value of water, or in addition transfer the bone volume from the real CT. We show an improved performance compared to both baseline pCTs suggesting that our method may be useful for MRI-only radiotherapy.