Part III | Chapter 11 216 reduce editing times, as shown in chapter 3. The ‘EVolution’ algorithm proposed and used in chapter 3 is not considered the ultimate DIR algorithm, but it had previously been shown to deliver clinically acceptable results. The EVolution algorithm employs global regularisation and thus ignores the fact that different tissues have different properties. For example, bladder and rectal tissue have different elastic properties compared to normal and cancerous prostate tissue. Ignoring these different properties can lead to implausible deformations in certain tissues. Newer, more complex DIR algorithms, such as ‘adaptive-EVolution’ (AEVO) as proposed by Zachiu et al.20, take into account differences in tissue properties by applying local regularisation. Evaluation of AEVO in prostate cancer patients treated on an MR-Linac showed comparable results to EVolution with respect to contour propagation, but much improved results in terms of anatomically plausible deformations within elastic tissues.20 With computational times of under 10 sec for mono-modal MRI-MRI registration, this algorithm is still suitable for a fast online adaptive workflow. As mentioned in chapter 3, we did not perform an extensive comparison of different methods for intrafraction contour propagation/generation. Besides more traditional segmentation algorithms, there are other ways to obtain (accurate) contours that can clinically be used, either directly or after minimal manual editing.21 This includes deep learning (DL)-based auto-segmentation or deformable contour propagation using convolutional neural networks.22,23 DL-based segmentation has already been implemented for offline contour segmentation for, among other indications, prostate cancer, where it has shown enormous potential.22 The output of DL-based networks relies heavily on the quality and amount of input data.21 Given the MR-Linac produces large amounts of clinical, imaging, and contouring data, infrastructures gathering this data (such as the MOMENTUM study, NCT0407530524) provide the ideal setting to train these networks. Nevertheless, the presence of contouring variability between the datasets will remain a challenge and therefore international, standardised contouring guidelines are required. Not only the quality of the output (contours) is important, but also other uses of image registrations algorithms should be considered when deciding which one to use. For example, the deformation vector fields that are obtained through DIR algorithms can be used for dose accumulation purposes.25 In the end, from all the available options, the option that best fits our technical and clinical goals should be selected and implemented in a clinically available workflow. To achieve this, the technical and clinical goals should be clearly defined, and methods should be tested in a systematic and comparable way. Current treatment planning times (component 3) for prostate cancer SBRT are relatively long for ATS workflows, in which a new treatment plan is completely re-optimised using the updated anatomy and daily contours.8,14 The treatment planning algorithm must take into account the Lorentz force on secondary electrons, for instance in the electron return effect.26 Currently, a graphics processing unit (GPU)-based Monte Carlo dose engine is commercially available and used for this purpose, but this is still relatively slow, with reported plan optimisation times of 5.5 min for prostate cancer.14,27 Faster treatment plan optimisation may reduce the time between imaging and actual treatment delivery and thereby could limit intrafraction motion.28 New and improved optimisation algorithms are currently investigated that could significantly reduce the optimisation times while providing similar treatment plan quality.28,29 These innovations, again, will potentially minimise the effects of intrafraction motion and deformations during treatment re-planning and could thereby increase
RkJQdWJsaXNoZXIy MTk4NDMw