ESTRO 2020 Abstract book

S939 ESTRO 2020

In MR-guided radiotherapy, contours of the target are propagated for each fraction to perform online daily adaptations. Conventional methods use deformable image registration which is often too slow for online use during treatment sessions. In this study, we investigate fast organ contour propagation from pre-treatment to fraction images in MR-guided prostate radiotherapy with convolutional neural networks (CNNs) that are trained for combined image registration and contour propagation. Material and Methods Five prostate cancer patients underwent twenty fractions of image-guided external beam radiotherapy on a 1.5T MR- Linac system (Unity, Elekta AB). For each patient, a pre- treatment 3D T2-weighted TSE MRI was used to delineate the CTV. The same scan was repeated during each fraction, with CTV contours being manually adapted if necessary. A 3D-CNN was trained in a supervised manner for combined image registration and contour propagation on a synthetically generated ground truth of randomly deformed images and prostate contours. The network estimated the propagated contour and a deformation field between the two input images (Fig1). The training set consisted of a synthetically generated ground truth of 15,000 randomly deformed images and prostate contours. We performed a leave-one-out cross-validation on the five patients and propagated the contours from the pre- treatment to the fraction scans. Three variants of the CNN, aimed at investigating supervision based on maximizing contour overlap, minimizing the error in the deformation vector fields over the field-of-view and, and a combination of the two were compared to a previously validated deformable registration approach (Elastix), by computing contour overlap, Hausdorff distances, and the prostate centroid distance for each of the propagated contours.

Conclusion Convolutional neural networks trained on the combined objective of maximizing prostate overlap and minimizing registration errors or on contour overlap only resulted in the best performance against networks trained on minimizing registration errors only. Convolutional networks are significantly faster than conventional DIR methods while achieving superior accuracy. Obtaining deformation fields and contours within a second is essential to facilitate intra-fraction adaptations. PO-1625 Deformable image registration (DIR) and radiobiological recalculation for retreat plan evaluation L. Stenzel 1 , R. Flynn 2 , M. Moore 3 1 National University of Ireland Galway, Dept. of Physics, Galway, Ireland ; 2 Mid-Western Radiation Oncology Centre- University Hospital Limerick, Radiotherapy Physics, Dooradoyle- Limerick, Ireland ; 3 University Hospital Galway, Medical Physics & Clinical Engineering, Galway, Ireland Purpose or Objective The study aimed to evaluate the application of DIR and radiobiological dose calculations when applied to multiple treatment plans for patients requiring retreatment. Retreatment was defined as patients who had undergone previous radiotherapy treatment and who returned for further treatment due to recurrent cancer or tumour growth localised near a previously treated area. To evaluate the potential clinical significance of DIR and radiobiological recalculation in retreat plan assessment, three methods of dose accumulation were investigated during this study. Material and Methods DIR and dose accumulation analysis were carried out on six patient retreat cases which involved non-standard fractionation. Dose accumulated in specific OARs was investigated. The dose accumulation methodologies varied in terms of image registration and dose calculation: (a) rigid physical accumulation (b) deformable physical accumulation

Results The neural networks trained on contour overlap or the combined objective achieved significantly better Hausdorff distances between predicted and ground truth contours than Elastix, with an additional much shorter time duration of 0.5 seconds ( Tab 1, Fig2).