ESTRO 2024 - Abstract Book

S3002

Physics - Autosegmentation

ESTRO 2024

[1] Nash et al, 2022. Physica Medica, 100, 112-119.

[2] Brock et al, 2017. Medical Physics, 44, 7, e43-e76.

330

Proffered Paper

External validation of rectal cancer CTV auto-contouring with patient-specific prior for online ART

Nicole Ferreira Silverio 1 , Wouter van den Wollenberg 1 , Anja Betgen 1 , Lisa Wiersema 1 , Corrie Marijnen 1 , Femke Peters 1 , Uulke A. van der Heide 1 , Rita Simoes 1 , Erik van der Bijl 2 , Martijn Intven 3 , Tomas Janssen 1 1 The Netherlands Cancer Institute, Department of Radiation Oncology, Amsterdam, Netherlands. 2 Radboud University Medical Centre, Department of Radiation Oncology, Nijmegen, Netherlands. 3 University Medical Center Utrecht, Department of Radiation Oncology, Utrecht, Netherlands

Purpose/Objective:

Deep learning has proven to be beneficial for auto-contouring in radiotherapy to speed up the process and reduce workload. However, segmentation models for clinical target volumes (CTV) where clinical interpretation plays a pivotal role, have shown less success. One of the reasons is that clinical interpretation varies between institutions and clinicians. Besides, variation in CTV segmentation is also caused by anatomical variations among (and within) patients as well as tumor dependent factors. These patient-specific factors hamper the generalizability of deep learning models that rely on medical images as input only. We hypothesize that, when considering daily re-contouring of the CTV for online adaptive radiotherapy, this problem can be mitigated, since the delineation made in the preparation phase implicitly contains the patient-specific clinical assessment. In this work we study the value of using this delineation as prior knowledge in multi-center generalizability of an auto-contouring model for the mesorectum CTV, to be used for online adaptive MR-guided radiotherapy of rectal cancer. In this study, 3D T2 weighted MRI scans and accompanying manual mesorectum delineations of 30 intermediate risk or locally advanced rectal cancer patients (483 scans) treated in-house on a 1.5T Unity MR-Linac (Elekta AB) were used for model training. The data was split on the patient level into training-validation-test sets (20-5-5 patients, respectively). To obtain the patient-specific prior containing the pre-treatment clinical assessment, per patient the mesorectum delineation on the planning scan was combined with population-based probability of likely inter-fraction mesorectum deformations. These probabilities were extracted from in-house mesorectum delineations on CBCT scans. The deformation probabilities were applied to the CTV delineation on the planning scan and thus provided a 3D voxel-wise patient-specific probability map of likely CTV delineation on subsequent imaging. This map was then registered rigidly to each daily scan of the patient. Two models were developed using the nnU-Net framework: a model without probability maps, trained only on MR-images, and a model with the probability map as an additional patient-specific input channel. We designate these models as MR-only and MR+prior. Material/Methods:

For the multicenter validation of the trained models, data was retrieved via the Momentum registry of rectal cancer patients from two external centers with MR-guided radiotherapy; center 1 (66 patients, 362 scans) and center 2 (23