Comparing the Performance of Radiation Oncologists versus a Deep Learning Dose Predictor to Estimate Dosimetric Impact of Segmentation Variations for Radiotherapy

Amith Jagannath Kamath, Zahira Mercado Auf der Maur, Robert Poel, Jonas Willmann, Ekin Ermis, Elena Riggenbach, Nicolaus Andratschke, Mauricio Reyes

Keywords: segmentation, clinical impact, radiation dose, validation

Abstract: Current evaluation methods for quality control of manual/automated tumor and organs-at- risk segmentation for radiotherapy are driven mostly by geometric correctness. It is however known that geometry-driven segmentation quality metrics cannot characterize potentially detrimental dosimetric effects of sub-optimal tumor segmentation. In this work, we build on prior studies proposing deep learning-based dose prediction models to extend its use for the task of contour quality evaluation of brain tumor treatment planning. Using a test set of 54 contour variants and their corresponding dose plans, we show that our model can be used to dosimetrically assess the quality of contours and can outperform clinical expert radiation oncologists while estimating sub-optimal situations. We compare results against three such experts and demonstrate improved accuracy in addition to time savings. Our code is available at https://github.com/ubern-mia/radonc-vs-dldp.

Submission Number: 45