Fast and accurate contour propagation 57 manual adaptation was needed in some cases, it could probably be performed within 3 min in the far majority of the fractions. Online adaptive radiotherapy workflows come with specific needs in terms of DIR technology. Algorithms need to be fast, accurate, and easy to use for the operator. While there are many registration algorithms available in the literature, very few fulfil these requirements, and even fewer have been validated for clinical use. For this work, we selected EVolution based on its demonstrated accurate performance for MR-to-MR contour propagation.18,20 The results obtained in the current study are in good correspondence with previous reports, since EVolution delivered overall clinically usable propagated contours. This was particularly the case for instances in which the time interval between sequential MR scans was shorter. In these cases, gradual volume changes and translations, due to bladder filling or drifts of the prostate were less extreme.1 Our results thus suggest that short cycle times (times between two MR images) are an important factor in the clinical accuracy of intrafraction propagated contours, and they should therefore be kept as short as possible. The main source of inaccuracies stemmed frommajor deformations occurring within the rectum, for example in case of a large gas pocket. In such instances, we hypothesise that the large magnitude of the deformations together with the significantly different image features introduced by the gas pocket itself has led to the algorithm converging towards a local minimum and in turn causing a local misregistration. Our previous work on intrafraction motion indicated that these rectal deformations are unpredictable and non-gradual.1,14,22,23 Especially cases with large rectal deformations could benefit from an adaptive workflow and therefore warrant extra time to assure contours are accurate. In terms of computational time, the algorithm converged in approximately 1.5–2.0 sec, which ensures smooth progress of online adaptive workflows that are as of now already time consuming (approximately 45 min per fraction for prostate cancer1). Furthermore, the algorithm’s control parameters were maintained at fixed values for all registered MR pairs. Once the algorithm has been configured for registering MR images acquired using a particular acquisition sequence, the same configuration can be maintained for any number of registered image pairs.19 This is beneficial for online adaptive workflows on an MR-Linac since there is no requirement for online tuning of algorithm parameters. Therefore, EVolution generally fulfils the technical and functional requirements for clinical use in a VF workflow. This paper is inherently limited by the exploratory design. We did not carry out a full comparison of e.g. different DIR algorithms or other auto-contouring solutions. Our aim was to assess the clinical quality of the contours provided by EVolution, so that it can serve as a basis for our future work regarding intrafraction adaptive workflows, and not to identify the most accurate auto-contouring solution. We only presented results for mono-modal MR-MR registration, since the intended use is for an MR-only MR-Linac workflow. As presented previously, this generally leads to better results in terms of Dice’s similarity coefficient compared to CT-MR or multi-model MR-MR registration.18 The results are therefore not applicable to multi-modal image registration. Additionally, only subjective assessments of the contours were conducted. Nevertheless, agreement rates were high for CTV and bladder contours, which mostly needed no or only minor editing (Table 1). We believe that the 3
RkJQdWJsaXNoZXIy MTk4NDMw