Owing to its excellent soft-tissue contrast, magnetic resonance (MR) imaging has found an increased application in radiation therapy (RT). By harnessing these properties for treatment planning, automated segmentation methods can alleviate the manual workload burden to the clinical workflow. We investigated atlas-based segmentation methods of organs at risk (OARs) in the head and neck (H&N) region using one approach that selected the most similar atlas from a library of segmented images and two multi-atlas approaches. The latter were based on weighted majority voting and an iterative atlas-fusion approach called STEPS. We built the atlas library from pre-treatment T1-weighted MR images of 12 patients with manual contours of the parotids, spinal cord and mandible, delineated by a clinician. Following a leave-one-out cross-validation strategy, we measured the geometric accuracy by calculating Dice similarity coefficients (DSC), standard and 95% Hausdorff distances (HD and HD95), and the mean surface distance (MSD), whereby the manual contours served as the gold standard. To benchmark the algorithm, we determined the inter-observer variability (IOV) between three observers. To investigate the dosimetric effect of segmentation inaccuracies, we implemented an auto-planning strategy within the treatment planning system Monaco (Elekta AB, Stockholm, Sweden). For each set of auto-segmented OARs, we generated a plan for a 9-beam step and shoot intensity modulated RT treatment, designed according to our institution's clinical H&N protocol. Superimposing the dose distributions on the gold standard OARs, we calculated dose differences to OARs caused by delineation differences between auto-segmented and gold standard OARs. We investigated the correlations between geometric and dosimetric differences. The mean DSC was larger than 0.8 and the mean MSD smaller than 2 mm for the multi-atlas approaches, resulting in a geometric accuracy comparable to previously published results and within the range of the IOV. While dosimetric differences could be as large as 23% of the clinical goal, treatment plans fulfilled all imposed clinical goals for the gold standard OARs. Correlations between geometric and dosimetric measures were low with R < 0.5. The geometric accuracy and the ability to achieve clinically acceptable treatment plans indicate the suitability of using atlas-based contours for RT treatment planning purposes. The low correlations between geometric and dosimetric measures suggest that geometric measures alone are not sufficient to predict the dosimetric impact of segmentation inaccuracies on treatment planning for the data utilised in this study.

Online adaptive radiotherapy would greatly benefit from the development of reliable auto-segmentation algorithms for organs-at-risk and radiation targets. Current practice of manual segmentation is subjective and time-consuming. While deep learning-based algorithms offer ample opportunities to solve this problem, they typically require large datasets. However, medical imaging data is generally sparse, in particular annotated MR images for radiotherapy. In this study, we developed a method to exploit the wealth of publicly available, annotated CT images to generate synthetic MR images, which could then be used to train a convolutional neural network (CNN) to segment the parotid glands on MR images of head and neck cancer patients. Methods: Imaging data comprised 202 annotated CT and 27 annotated MR images. The unpaired CT and MR images were fed into a 2D CycleGAN network to generate synthetic MR images from the CT images. Annotations of axial slices of the synthetic images were generated by propagating the CT contours. These were then used to train a 2D CNN. We assessed the segmentation accuracy using the real MR images as test dataset. The accuracy was quantified with the 3D Dice similarity coefficient (DSC), Hausdorff distance (HD) and mean surface distance (MSD) between manual and auto-generated contours. We benchmarked the approach by a comparison to the inter-observer variation determined for the real MR images, as well as to the accuracy when training the 2D CNN to segment the CT images. Results: The determined accuracy (DSC: 0.77±0.07, HD: 18.04±12.59 mm, MSD: 2.51 ± 1.47 mm) was close to the inter-observer variation (DSC: 0.84±0.06, HD: 10.85±5.74 mm, MSD: 1.50±0.77 mm), as well as to the accuracy when training the 2D CNN to segment the CT images (DSC: 0.81±0.07, HD: 13.00±7.61 mm, MSD: 1.87±0.84 mm). Conclusion: The introduced cross-modality learning technique can be of great value for segmentation problems with sparse training data. We anticipate using this method with any non-annotated MRI dataset to generate annotated synthetic MR images of the same type via image style transfer from annotated CT images. Furthermore, as this technique allows for fast adaptation of annotated datasets from one imaging modality to another, it could prove useful for translating between large varieties of MRI coni Accepted Article This article has been accepted for publication and undergone full peer review but has not been through the copyediting, typesetting, pagination and proofreading process, which may lead to differences between this version and the Version of Record. Please cite this article as

Adaptive online MRI-guided radiotherapy of head and neck cancer requires the reliable segmentation of the parotid glands as important organs at risk in clinically acceptable time frames. This can hardly be achieved by manual contouring. We therefore designed deep learning-based algorithms which automatically perform this task.Imaging data comprised two datasets: 27 patient MR images (T1-weighted and T2-weighted) and eight healthy volunteer MR images (T2-weighted), together with manually drawn contours by an expert. We used four different convolutional neural network (CNN) designs that each processed the data differently, varying the dimensionality of the input. We assessed the segmentation accuracy calculating the Dice similarity coefficient (DSC), Hausdorff distance (HD) and mean surface distance (MSD) between manual and auto-generated contours. We benchmarked the developed methods by comparing to the inter-observer variability and to atlas-based segmentation. Additionally, we assessed the generalisability, strengths and limitations of deep learning-based compared to atlas-based methods in the independent volunteer test dataset.With a mean DSC of 0.85± 0.11 and mean MSD of 1.82 ±1.94 mm, a 2D CNN could achieve an accuracy comparable to that of an atlas-based method (DSC: 0.85 ±0.05, MSD: 1.67 ±1.21 mm) and the inter-observer variability (DSC: 0.84 ±0.06, MSD: 1.50 ±0.77 mm) but considerably faster (<1s v.s. 45 min). Adding information (adjacent slices, fully 3D or multi-modality) did not further improve the accuracy. With additional preprocessing steps, the 2D CNN was able to generalise well for the fully independent volunteer dataset (DSC: 0.79 ±0.10, MSD: 1.72 ±0.96 mm)We demonstrated the enormous potential for the application of CNNs to segment the parotid glands for online MRI-guided radiotherapy. The short computation times render deep learning-based methods suitable for online treatment planning workflows.