A Multi-criteria Quality Assessment of Automated Alternative Segmentations for Radiation Therapy of Brain Tumor Patients

Amith Jagannath Kamath, Roberto Münger, Robert Poel, Elias Rüfenacht, Alain Jungo, Jonas Willmann, Nicolaus Andratschke, Mauricio Reyes

Keywords: Radiotherapy Quality Assurance, Alternative Segmentations, Clinical Evaluation.

TL;DR: This paper describes a Quality-Assurance workflow using Stochastic Segmentation Network (SSN) and Gaussian Process Sampling Segmentation of Images (GPSSI) to perform a multi-criteria assessment of segmentations pertaining to Glioblastoma.

Abstract: Medical image segmentation is a crucial part of the radiotherapy (RT) planning workflow for treating brain tumor patients. Consequently, radiotherapy quality assurance (RTQA) of expert-segmentations is performed in clinical routine, especially in clinical trials. However, RTQA is time-consuming and error-prone. Towards automating this, we hypothesize that models that can generate a distribution of segmentations to simulate the variability in expert-segmentation, can be used as a proxy to evaluate the compliance and quality of new segmentations. In this paper, we evaluate a deep learning (Stochastic Segmentation Networks), and a non-deep learning approach (Gaussian Process Sampling Segmentation of Images) to generate this distribution of ‘alternative segmentations’. We assess the quality of these alternative segmentations using three complementary criteria: (i) a Turing-Test like review comparing expert-segmentations with computer-generated alternatives, (ii) geometric compliance using Dice similarity coefficient, and a (iii) dosimetric compliance evaluation using dose-volume histogram curves. On an evaluation data set consisting of 40 RT plans, our results indicate that these methods yield plausible alternative segmentations which could be used to build a deep learning-based RTQA platform. Our results further indicate that geometric compliance needs to be complemented with dosimetry to fully characterize the impact of segmentation deviations for target coverage and organs at risk toxicity.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: validation/application paper

Primary Subject Area: Application: Other

Secondary Subject Area: Validation Study

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.