Automatic Identification of Segmentation Errors for Radiotherapy Using Geometric Learning

Automatic segmentation of organs-at-risk (OARs) in CT scans using convolutional neural networks (CNNs) is being introduced into the radiotherapy workflow. However, these segmentations still require manual editing and approval by clinicians prior to clinical use, which can be time consuming. The aim this work was develop a tool automatically identify errors 3D OAR without ground truth. Our uses novel architecture combining CNN graph network (GNN) leverage segmentation’s appearance shape. proposed model trained data-efficient learning synthetically-generated dataset parotid gland with realistic contouring errors. effectiveness our assessed ablation tests, evaluating efficacy different portions as well use transfer from custom pretext task. best performing predicted on precision 85.0% & 89.7% for internal external respectively, recall 66.5% 68.6%. This offline QA could used pathway, potentially decreasing spend correcting contours detecting regions their attention. All code publicly available at https://github.com/rrr-uom-projects/contour_auto_QATool .