ESTRO 2024 - Abstract Book

S4312

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

ESTRO 2024

2143

Poster Discussion

Intra-fraction Tumour and Multi-Organ-At-Risk Tracking for Pancreatic Cancer SABR

Abdella M Ahmed 1 , Levi Madden 1 , Maegan Stewart 1 , Adam Mylonas 2 , Ryan Brown 1 , John Kipritidis 1 , Danielle Crystall 1 , Robert Finnegan 1 , Andrew Kneebone 1 , George Hruby 1 , Paul Keall 2 , Jeremey Booth 1 1 Northern Sydney Cancer Centre, Radiation Oncology, Sydney, Australia. 2 Image-X Institute, School of Health Sciences, University of Sydney, Sydney, Australia

Purpose/Objective:

Stereotactic ablative radiotherapy (SABR) of the pancreas is characterised by high doses per fraction, and sharp dose gradients to spare organs-at-risk (OARs). Intra-fraction motion of the pancreas and OARs can be large and complex with anatomical deformation and respiratory motion, requiring compromises with margins, coverage and dose levels, even for gated delivery. Intra-treatment monitoring on a standard linear accelerator is limited to surface imaging and kV/MV imaging of fiducials near the lesion. A more sophisticated approach to track the target and OARs is required to maximise the efficacy of treatment. Deep learning-based methods have shown success in tracking tumour motion in intra-fraction kV images. However, combined tumour and OAR tracking has not yet been performed. To fill this knowledge gap, we explore a deep learning approach utilising kilovoltage (kV) images to achieve the task of intra fraction markerless tumour and OAR segmentation for motion tracking during pancreatic cancer SABR.

Material/Methods:

Patient data from an ethics approved trial for pancreatic cancer SABR was utilised to train and test a conditional generative adversarial network (cGAN) model for segmentation of gross tumour volume (GTV) and OARs (pancreas head, pancreas, and duodenum) in intra-fraction planar kV imaging. The cGAN architecture had a UNet generator and a patch-GAN classifier. Labelled digitally reconstructed radiographs (DRRs) were generated using an in-house algorithm, generated from contoured planning CT (DRR pCT ) and CBCT (DRR CBCT ) image sets acquired in exhale breath hold. A population model was trained using DRR pCT from six patients (21,600 DRRs). Patient-specific models were created on a fraction-by-fraction basis by fine-tuning the population model using corresponding DRR CBCT for training (360 DRRs per fraction). The model was evaluated on intra-fraction kV images (898 kV images obtained from 14 fractions of 3 patients), by comparing the predicted segmentation to the forward projection of the pre-treatment CBCT contours, after couch-shift correction is applied. The centroid difference along the x- and y- direction, the Dice similarity coefficient (DSC), and the 95th percentile of the Hausdorff-distance (HD95) were calculated. The mean ± 1SD of all the metrics were reported.

Results:

The full result for each organ tracking is given in Table 1. For all organs, the centroid distance in x- and y- directions were on average less than 3.3 mm and 1.2 mm, respectively. The mean DSC and HD95 was above 0.92 and under 8.2 mm, respectively. An example for training, organ prediction and simultaneous tumour and OAR tracking is shown in Figure 1. The computational time to fine-tune the population model to a patient specific model on a GPU (NVIDIA