ESTRO 2024 - Abstract Book

S3060

Physics - Autosegmentation

ESTRO 2024

1501

Proffered Paper

Deep learning-based contrast agent segmentation for liver cancer targeting during radiation therapy

Adam Mylonas 1 , Doan Trang Nguyen 1 , Venkatakrishnan Seshadri 2 , Prabhakar Ramachandran 2 , Jessica Lye 3 , Richard Khor 3 , Mark Gardner 1 , Chandrima Sengupta 1 , Marco Mueller 1 , Paul Keall 1 1 The University of Sydney, Image X Institute, Sydney, Australia. 2 Princess Alexandra Hospital, Medical Physics, Brisbane, Australia. 3 Olivia Newton John Cancer Wellness & Research Centre, Department of Radiation Oncology, Melbourne, Australia

Purpose/Objective:

Liver cancer is the sixth most common cancer globally and the third leading cause of cancer-related deaths. 1 One of the challenges of radiation therapy is the significant movement of tumours that can occur during treatment, particularly for tumours in the liver. This motion can potentially lead to cancer target underdose and healthy tissue overdose, compromising treatment efficacy. 2 Real-time image-guided adaptive radiation therapy can track the target and account for the motion. 3 Typically, fiducial markers are surgically implanted as a surrogate of the tumour position due to the low radiographic contrast of the soft tissues in kilovoltage (kV) images. However, there are costs, time delays, and surgical complications associated with marker insertion. Patients being treated for hepatocellular carcinoma, the most common form of primary liver cancer, often receive chemotherapy via transcatheter arterial chemoemobilisation (TACE). A radio opaque contrast agent is injected during TACE and remains visible on imaging for months afterwards. The contrast agent could be used as a surrogate of the tumour, enabling a non-invasive method for real time tumour tracking. 4 Therefore, the goal of this work was to develop and investigate a deep learning-based contrast agent segmentation for real-time liver cancer targeting during radiation therapy. A conditional generative adversarial network (cGAN) was developed to segment the contrast agent in 2D kV images. The model leverages the patient’s pretreatment imaging and planning data that is available prior to the commencement of their treatment. The patient-specific model can be incorporated into the treatment workflow, as shown in Fig. 1. The model was evaluated on an imaging database of three liver cancer patients from the ROCK-RT trial (NCT05169177). 5 A patient-specific model was trained for each patient using 36,000 digitally reconstructed radiographs (DRR) per patient. The DRRs were produced using the 3D planning CT and contrast agent contour volumes. The volumes were forward projected to produce one digitally reconstructed radiograph (DRR) each 1 degree over 360 degrees, generating 360 projections. Prior to training, data augmentation for each patient was performed 100 times by randomly shifting the CT geometry up to 10 mm and rotating up to 10 degrees and then computing a new set of DRRs. In total 36,000 DRRs were created for each patient. A 70 × 70 PatchGAN was used for the discriminator architecture and 256 × 256 UNet for the generator architecture. 6 Each model was trained for two epochs with a batch size of one and a learning rate of 0.0002 using the Adam optimiser. Material/Methods: