# Automatic identification of segmentation errors for radiotherapy using geometric learning

**Authors:** Edward G. A. Henderson, Andrew F. Green, Marcel van Herk, Eliana M. Vasquez Osorio  
**arXiv:** [2206.13317](https://arxiv.org/abs/2206.13317)  
**Date:** 2022-06-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.13317v1/

## Key points

- Convolutional neural networks (CNNs) are used for automatic segmentation of organs-at-risk (OARs) in CT scans
- Manual editing and approval by clinicians is still required prior to clinical use, which can be time-consuming and prone to errors
- A tool was developed that automatically identifies errors in 3D OAR segmentations without the need for ground truth data
- The proposed tool uses a novel architecture that combines a CNN and graph neural network (GNN) to leverage both the appearance and shape of the segmentation
- The model was trained using self-supervised learning on a synthetically-generated dataset of segmentations of the parotid gland with realistic contouring errors
- Results showed that their best performing model predicted errors on the parotid gland with high precision and recall rates
- This offline quality assurance tool could potentially decrease the time clinicians spend correcting contours by detecting regions that require their attention
- The code used in this study is publicly available on GitHub for further development and implementation in clinical settings.

## Summary

The use of convolutional neural networks (CNNs) for automatic segmentation of organs-at-risk (OARs) in CT scans has become increasingly popular in radiotherapy workflows. However, these segmentations still require manual editing and approval by clinicians prior to clinical use, which can be time-consuming and prone to errors. To address this issue, a team of researchers from the University of Manchester and The Christie NHS Foundation Trust developed a tool that automatically identifies errors in 3D OAR segmentations without the need for ground truth data. The proposed tool uses a novel architecture that combines a CNN and graph neural network (GNN) to leverage both the appearance and shape of the segmentation. The model was trained using self-supervised learning on a synthetically-generated dataset of segmentations of the parotid gland with realistic contouring errors. Ablation tests were conducted to evaluate the efficacy of different portions of the architecture as well as transfer learning from an unsupervised pretext task. The results showed that their best performing model predicted errors on the parotid gland with a precision of 85.0% & 89.7% for internal and external errors respectively, and recall of 66.5% & 68.6%. This offline quality assurance tool could potentially decrease the time clinicians spend correcting contours by detecting regions that require their attention. This research paper has been accepted for presentation at the 25th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022). It is important to note that this preprint has not undergone peer review or any post-submission improvements or corrections yet. Overall, this study presents a promising approach towards improving radiotherapy workflows by automating error detection in OAR segmentations without relying on ground truth data. The code used in this study is publicly available on GitHub for further development and implementation in clinical settings. This could help reduce time spent manually editing OAR segmentations while increasing accuracy and reliability in radiotherapy treatments.