ESTRO 2020 Abstract book

S992 ESTRO 2020

Conclusion Volumetric 4D-MRI enabled positional probability mapping of organ respiratory motion. Positional probability map derived from volumetric 4D-MRI may hold potentials for respiratory motion gating and tracking in the future MRgRT. PO-1704 Global and local evaluation of deep learning contouring of prostate CT E. Brunenberg 1 , A. Lazareva 2 , I. Steinseifer 1 , R.J. Smeenk 1 , P. Aljabar 2 , R. Monshouwer 1 1 Radboud university medical center, Radiation Oncology, Nijmegen, The Netherlands ; 2 Mirada Medical Limited, Science Group, Oxford, United Kingdom Purpose or Objective As a proof of principle, deep learning contouring (DLC) of prostate CT has been shown to outperform atlas-based segmentation with respect to contouring accuracy [1]. However, larger training and test sets were required in order to increase confidence in the deep learning model before clinical implementation. This study compares an extended DLC model with manual ground truth, in addition taking into account local spatial information. Material and Methods An expert radiation oncologist delineated clinical contours on CT images of more than 300 patients that underwent primary prostate irradiation. Delineated ROIs comprised the prostate, seminal vesicles, bladder, anal and rectal inner walls (mostly defined by the used endorectal balloon), and femoral heads [2]. Two DLC models (WorkflowBox 2.0.1, Mirada Medical) were compared, one initially trained on 114 datasets for 29-52 iterations (depending on the ROI) [1], and the second further extended with 228 new datasets for 73-80 iterations. An independent test set of 52 patients was generated to assess the accuracy of the two DLC models against clinical contours (considered ground truth), using the Dice similarity coefficient (DSC). For the extended model also a binned (volume-normalized) DSC was calculated for four bins in superoinferior direction, in order to obtain more spatial information on model performance. All comparisons were tested for statistical significance (assuming equality as the null hypothesis). Results The boxplots in Figure 1 show global DSC median values and IQR for the two DLC models. At a confidence level of 95%, the null hypothesis of equality could be rejected for prostate, bladder and rectal wall contours. Figure 2 shows the binned DSC for the extended model, for the first four ROIs. Equality (compared to the global DSC) could be rejected for all bins, except for the vesicles’ middle bins. Contour overlap in the caudal bins is thus significantly less for all four ROIs, and also in the cranial bins for prostate, vesicles and bladder.

Conclusion This study has shown that extending the DLC training set with consistent data can substantially improve segmentation results. In addition, the local information provided by the binned DSC is very useful. The high variability at caudal and cranial ends can have various causes: deviations from the delineation protocol, uncertainties in anatomy at organ borders, or partial volume effects. Therefore, besides numerical DSC results, clinical evaluation of DLC results is an important next step. References 1. Aljabar P, et al. OC-0419 Comparison of auto-contouring methods for regions of interest in prostate CT. R&O 2018; 127(Suppl 1), pp. S218-9. 2. Smeenk RJ, et al. Dose-effect relationships for individual pelvic floor muscles and anorectal complaints after prostate radiotherapy. IJROBP 2012; 83(2), pp. 636- 44. PO-1705 Attention boosted convolution neural networks for parotid and parotid tumor segmentation on MR image W. Jiazhou 1 , H. Weigang 1 , Y. Fangfang 2 , X. Xianwu 2 1 Fudan University Shanghai Cancer Center, Radiotherapy Department, Shanghai, China ; 2 Municipal Hospital of Taizhou University School of Medicine, Oncology Depearment, Taizhou, China Purpose or Objective To develop a deep learning-based auto segmentation method for parotid and parotid tumor segmentation on MR