ESTRO 2024 - Abstract Book

S3784

Physics - Image acquisition and processing

ESTRO 2024

604

Poster Discussion

Towards clinical sustainability of deep learning-based sCT models after CBCT image software upgrade

Arthur Jr V Galapon, Dirk Wagenaar, Johannes Albertus Langendijk, Stefan Both

University Medical Center Groningen, Radiotherapy, Groningen, Netherlands

Purpose/Objective:

Deep learning-based synthetic CTs (sCTs) can be generated directly from CBCT, improving image quality and enabling dose calculation, as demonstrated in various studies [1-3]. In clinical practice, changes in CBCT image quality frequently occur due to hardware and software system upgrades, which may require validation or retraining of the sCT-generating models. Significant changes in CBCT image quality can lead to a decline in the quality of generated sCTs, as these models often rely on the images used during training. Moreover, workflows dependent on sCTs may be disrupted while centers acquire new datasets based on post-upgrade images for model retraining. In this study, we propose and test a methodology for the rapid re-validation of sCT-generating networks after a CBCT software upgrade using transfer learning. We also evaluated the minimum number of images required to achieve the same performance as an existing accepted synthetic CT model.

Material/Methods:

Paired verification CT (rCT) and CBCT images of 70 head and neck cancer patients acquired after a scheduled software upgrade of the onboard CBCT scanner (IBA Proteus 235 P+) were used. An initial deep convolutional neural network (DCNN) sCT model [1] was trained using images obtained before the upgrade was used as the baseline model. Transfer-learning models (TL-models) were trained using a subset of 5, 10, 15, 20, 25, and 30 patients extracted from the primary database. Each subset was trained using a 70% training and 30% validation split ratio. A separate test set comprising ten patients was reserved to evaluate the performance of each model. During transfer learning, the existing DCNN model weights were retrained by freezing the old model’s feature-extracting (encoder) layers while updating all other layers’ weights. This method, which utilized existing trained weights, is expected to reduce the number of images required during re-training. Early stopping was implemented after the validation loss ceased to decrease after several epochs. Additionally, a new model was trained using images from 60 patients to compare the updated TL models’ performance with that of a newly trained model. The mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and 95% Hausdorff distance of bone [>250 HU] (HD95) were quantified to evaluate the sCT image quality. The rCT served as the ground truth. A pre-upgrade image quality metric, evaluated using the sCTs generated from the test set with the old model, was used to compare the pre-and post-upgrade quality of the sCTs. The mean doses of the target (CTV) and organs at risk (OARs), ΔNTCP values, and the gamma pass rates obtained from the clinical treatment plans were utilized to assess the dosimetric performance of the synthetic CTs. Moreover, the old model's image quality and dose calculation performance will serve as the baseline [1,3].