ESTRO 2024 - Abstract Book

S3023

Physics - Autosegmentation

ESTRO 2024

2. Alzahrani N, Henry A, Clark A, Murray L, Nix M, Al-Qaisieh B. Geometric evaluations of CT and MRI based deep learning segmentation for brain OARs in radiotherapy. Phys Med Biol. 2023;68(17).

726

Mini-Oral

Weakly supervised commissioning of externally developed auto-segmentation models

Bastiaan W.K. Schipaanboord 1 , Peter J. Koopmans 2 , Erik van der Bijl 2 , Charlotte L. Brouwer 3 , Tomas Janssen 1

1 The Netherlands Cancer Institute, Department of Radiation Oncology, Amsterdam, Netherlands. 2 Radboud University Medical Center, Department of Radiation Oncology, Nijmegen, Netherlands. 3 University Medical Center Groningen, Department of Radiation Oncology, Groningen, Netherlands

Purpose/Objective:

In online adaptive radiotherapy, segmentation is performed under time pressure. This makes it an ideal setting to introduce auto-segmentation. However, high-quality segmentation data to commission the segmentation model on is scarce, since segmentation performed under time pressure is less consistent. When introducing an auto-segmentation model into clinical practice, one would like to assess both the quality of the predicted segmentations and the robustness of the model over a wide range of anatomical variation and/or image quality, representative of daily clinical practice. Evaluating the quality of the model is generally performed by comparing a number of randomly selected cases against manual ground-truth segmentations on various volume and distance metrics. Evaluating the robustness of the model, however, tends to be more difficult. Especially, when the model is provided by an external party (model supplier) and the institution introducing the model (receiving hospital) does not possess a high quality dataset to commission the model on. Assuming that a segmentation model is more likely to fail for an atypical case as opposed to a more average one, we propose a methodology that selects cases based on unsupervised anomaly detection. In this way, the selected cases are more representative for the variation observed in clinical practice than would be the case when selecting cases randomly. This therefore provides a more suitable set of cases to evaluate the robustness of the model on. For this approach, the receiving hospital does not require a high-quality dataset with well-curated segmentations. Instead, the model supplier provides a set of image/shape features of the data on which the model was trained, that correlate with model performance. Then the receiving hospital can use these features to train an unsupervised anomaly detector on a large dataset of unlabeled cases, e.g. all scans of the past year for a specific (online-adaptive) treatment, and use the anomaly scores to select a representative set for manual evaluation The anomaly detector is trained in an unsupervised fashion and thus does not require a large, high quality, curated dataset, yet it is primed with information (choice of features) from the training data. We refer to this methodology as weakly supervised commissioning.

Material/Methods: