As part of treatment planning for radiotherapy, the Organs at Risk (OARs) are delineated on the patient's CT scan. This work aims to develop a method to measure variability in OAR delineations and detect errors. Approach. A normative modelling approach was implemented by training a Variational Autoencoder (VAE) on a dataset of images and delineations to model the "acceptable" variability distribution. The trained VAE was then used to reconstruct unseen cases. Disagreements between input and reconstructed delineations highlighted regions where the input deviated from the training distribution. This approach was validated by evaluating the reconstructions of spinal cord and brainstem delineations where common clinical errors had been introduced. Main results. Results showed that the model successfully detected errors, even when only a few voxels or slices were added or removed. Distance to Agreement (DTA) maps were generated to quantify the magnitude of the disagreements in misclassified regions. These results were further validated by manually evaluating some of the test cases. Significance. This tool has the potential of assisting clinicians in reviewing and validating OAR delineations.
Dronne et al. (2026) studied this question.