ESTRO 2020 Abstract Book

S814 ESTRO 2020

Results The coverage of the primary and nodal target volumes was comparable for both techniques and for both subsets of patients. The primary planning target volume (PTV) receiving at least 95% of the prescription isodose ranged from 97.2±1.1% to 99.1±1.2% for H and from 96.5±1.9% to 98.3±0.9% for HT. For the nodal CTVs the dose received by 98% of the planning target volume ranged 55.5±0.1 to 56.0±0.8Gy for H and HT respectively. The only significant and potentially relevant differences were observed for the bowels. In this case V 40Gy resulted 226.3±35.9 and 186.9±115.9 cm 3 for the node positive and node negative patients respectively for Halcyon. The corresponding findings for HT were: 258.9±60.5 and 224.9±102.2cm 3 . On the contrary, V 15Gy resulted 1279.7±296.5 and 1557.2±359.9 cm 3 for HT and H respectively for node positive and 1010.8±3209 vs. 1203.8±332.8cm 3 Conclusion This study confirmed the equivalence between Halcyon based and Helical Tomotherapy based plans for the intensity modulated rotational treatment of cervix uteri cancer. Different levels of sparing were observed for the bowels with H better protecting in the high-dose region and HT in the mid-low dose regions. The clinical impact of these differences should be further addressed. PO-1508 Exploration of deformable image registration to augment training data for deep learning contouring R. Baggs 1 , P. Aljabar 1 , M. Gooding 1 , P. Poortmans 2 , Y. Kirova 2 1 Mirada Medical Limited, Science, Oxford, United Kingdom ; 2 Institut Curie, Radiation Oncology, Paris, France Purpose or Objective This work investigates whether deformable image registration of CT images, as a means for data augmentation, affects deep learning networks for auto- contouring structures in breast cancer patients. Augmentation can be used to improve convolutional neural networks for image segmentation, for example in [1], which deployed model-based augmentation for MRI, rather than CT, in a non-radiation therapy (RT) application. Material and Methods Simulation CT images were acquired for 50 patients undergoing RT for breast cancer. Manual contours of the breast and lymph node regions (axilla levels 1 to 4 and interpectoral) served as ground truth labels for evaluation. A Deep Learning Contouring (DLC) model was trained on 40 cases, with the remaining 10 split equally into cross- validation (CV) and test sets. The model was trained using jittering [2], i.e. applying random affine 2D transformations to training CT slices. During training, each batch of slices was augmented with the same number of jittered slices. The same 50 cases were then used to create a larger training set using offline deformable registration based on optical flow [Mirada RTx, Mirada Medical Ltd]. 451 registrations were randomly selected from the 2450 combinations, excluding the original CV/test data and gross registration errors. Incorporating the original 50, gave a total of 501 datasets. This final augmented dataset was used to train another DLC model, evaluating against the original CV and test sets to enable comparison. The accuracy of each method was measured using quantitative measures, the Dice similarity coefficient (DSC) and the median surface distance (SD). Results The DSC results showed that for each organ, the data augmentation method using deformable registration was comparable or better than the standard jittering method. Surface distances were comparable, or lower, for deformable registration showing a slight improvement in accuracy when using this method. The jittering method under-segmented the breast in two test cases, as seen in the figure’s outliers.

Figures: DSC scores and Median surface distance (SD) for both methods.

Conclusion The preliminary study shows that deformable image registration for data augmentation is comparable or slightly better than affine jittering. Due to the limited test set available, further investigation is needed with a larger test set to further explore the generalisabilty and robustness of the approach and to assess the extent of optimisation needed in training. Such an approach may allow training of high quality DLC models using fewer curated datasets. References [1] Lin, A data augmentation approach to train fully convolutional networks for left ventricle segmentation. Magnetic Resonance Imaging 2019 [2] Simonyan, Very Deep Convolutional Networks for Large-Scale Image Recognition, Computer Science, 2014 PO-1509 Application of RapidPlan in the optimization of planning for patients with prostate cancer M. Raczkowski 1 , T. Siudziński 1 , M. Janiszewska 1 , A. Maciejczyk 2 1 Lower Silesian Oncology Center, Medical Physics Departmen, Wroclaw, Poland ; 2 Lower Silesian Oncology Center, Department of Radiotherapy, Wroclaw, Poland