Assessing artificial intelligence MRI autocontouring in Raystation and the AutoConfidence uncertainty model for brain radiotherapy

Abstract

Abstract: Background: In radiotherapy, deep learning autosegmentation (DL-AS) and automation of quality assurance (QA) have the potential to efficiently standardize and enhance the quality of contours.

Aim: To assess the performance of DL-AS in delineating organs-at-risk (OARs) in brain RT using the RayStation Treatment Planning System. Secondly, to build and test a novel artificial intelligence QA model called AutoConfidence (ACo).

Methods: Retrospective MRI and CT cases were randomly selected for training and testing. DL-AS models were evaluated from geometric and dosimetric perspectives, focusing on the impact of pre-training editing. The ACo model was evaluated using two sources of autosegmentation: internal autosegmentations (IAS) produced from the ACo generator and two external DL-AS with different qualities (high and low quality) produced from RayStation models.

Results: The edited DL-AS models generated more segmentations than the unedited models. Editing pituitary, orbits, optic nerves, lenses, and optic chiasm on MRI before training significantly improved at least one geometry metric. MRI-based DL-AS performed worse than CT-based in delineating the lacrimal gland, whereas the CT-based performed worse in delineating the optic chiasm. Except for the right orbit, when delineated using MRI models, the dosimetric statistical analysis revealed no superior model in terms of the dosimetric accuracy between the MR and CT DL-AS models. The number of patients where the clinical significance threshold was exceeded was higher for the optic chiasm D1% than for other OARs, for all models. ACo had excellent performance on both internal and external segmentations across all OARs (except lenses). Mathews Correlation Coefficient was higher on IAS and low-quality external segmentations than high-quality ones.

Conclusion: MRI DL-AS in RT may improve consistency, quality, and efficiency but requires careful editing of training contours. ACo was a reliable predictor of uncertainty and errors on DL-AS, demonstrating its potential as an independent, reference-free QA tool.