ESTRO 2021 Abstract Book

S210

ESTRO 2021

Conclusion This work has illustrated that the existing Elekta protocol in establishing beam steering error for FFF beams can lead to significant impacts on beam position and subsequently on patient plan quality without machine interlocks or even significant reported steering errors. A change in philosophy is required when considering FFF beam steering errors to focus on positional accuracy as opposed to asymmetry. OC-0303 Using spatial probability maps to highlight potential inaccuracies in deep learning based contours W. Verbakel 1 , W. van Rooij 1 , B. Slotman 1 , M. Dahele 1 1 Amsterdam University Medical Centers, Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective Organs-at-risk (OAR) contouring in radiotherapy is a largely manual task. It is time-consuming and prone to variation. Deep learning-based delineation (DLD) shows promise in terms of quality and speed, but does not perform perfectly, in part because of variation in delineation of the training data. Therefore manual checking of DLD is recommended. This is a limitation, for example, when introducing DLD into daily online adaptive radiotherapy (ART; due to time taken, and the need to perform checking tasks under pressure). There are currently no commercial radiotherapy tools to focus attention on the areas of greatest uncertainty within a DLD contour. Therefore, we explored the use of spatial probability maps (SPMs) for this task, using salivary gland segmentation as the paradigm. Materials and Methods Data consisted of 315/264 clinically contoured parotid/submandibular glands, without further curation. A 3D fully convolutional network was trained on 5/6 and tested on 1/6 of the datasets. Subsequently, SPMs were created using Monte Carlo dropout (MCD). The method was boosted by placing a Gaussian distribution over the model's parameters during sampling (MCD+GD). This provided 101 predictions to quantify the uncertainty, while a majority vote defined the delineated structure. MCD and MCD+GD were quantitatively compared by calculating uncertainty from the SPM voxel values, and the SPMs were visually inspected. Results The 97 test cases had dice similarity coefficients between 0.75 and 0.92, except for 3 cases (0.61, 0.62 and 0.71, all three having a high average uncertainty). The addition of the Gaussian distribution increased the method's ability to detect uncertainty. In general, this technique demonstrated uncertainty in areas that (1) have lower contrast, (2) are less consistently contoured by clinicians and (3) deviate from the anatomical norm. Generating the SPM for 1 gland took <2.5 seconds. The figure below shows an example of SPM for a parotid gland with dice similarity coefficient of 0.86, and how this could be presented in the clinic.