ESTRO 2025 - Abstract Book

S3251

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

Purpose/Objective: Pancreatic cancer is aggressive with a 5-year survival rate of just 5%. [1] The use of stereotactic body radiation therapy (SBRT) has shown promise, particularly with on-couch MRI guidance. [2,3] Accurate motion management is crucial, but limited to CBCT, fiducials and surface guidance for standard linear accelerators. Deep learning with intra-fraction kV imaging may overcome the image quality deficits with kV images to enable markerless tracking of the abdominal structures during treatment. We investigate a deep learning approach to predict the GTV, pancreas head and the whole-pancreas contours in intra-fraction kV images and quantify the accuracy of prediction. Material/Methods: Patient data from an ethics-approved trial for pancreatic SBRT (NCT03505229) was used to train and test a conditional-generative-adversarial-network (cGAN) to segment the GTV, pancreas-head and whole-pancreas in 2D kV images acquired intra-fraction as shown in Figure 1. A dataset including planning CT, pre-treatment CBCT, intra fraction kV images were acquired for 25 patients. All images were acquired with the patient in exhale-breath-hold or gated at exhale. DRRs were generated from contoured planning-CT and half-fan CBCT. All DRRs were labelled by projecting 3D contours. A population model was initially trained using labelled CT-DRRs from 19 patients. Patient specific models were created for 6 additional patients (27 fractions) by fine-tuning the population model using CBCT-DRRs for each predicted structure type. The patient-specific models were trained on a fraction-by-fraction basis and tested on triggered kV images acquired during treatment. The predicted contours were compared to rigid registrations of pre-CBCT contours to fiducials (referred as shifted-preCBCT contours) in each kV image. The performance of each model was evaluated using Dice-Similarity-Coefficient (DSC) and average Hausdorff Distance (AHD).

Results: The models predicted GTV, pancreas-head and whole-pancreas contours in intra-treatment kV images with DSC of 0.87±0.08, 0.89±0.08 and 0.86±0.10, respectively. The AHD for these structures were 0.8±0.8 mm, 0.9±0.9 mm, and 1.3±-1.1 mm respectively. Figure 2 shows sub-millimetre tracking for GTV, and pancreas head was achieved. The prediction time of a structure by the cGAN was 32.9 ms.