ESTRO 2025 - Abstract Book

S3216

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

ESTRO 2025

Purpose/Objective: Tumour motion during liver radiation therapy can compromise treatment accuracy, potentially leading to

1-5 . Real-time tumour tracking enables precise radiation delivery, improving safety and

suboptimal outcomes

efficacy 6-8 . However, poor soft tissue contrast in X-ray images and significant liver tumour motion makes this task challenging. Traditional modes of marker segmentation, such as template matching, report reduced accuracy when fiducials are obscured by bone, surgical clips, and stents. Moreover, they necessitate the generation of a predefined . Though deep-learning approaches have demonstrated potential in addressing these limitations, existing deep-learning based marker template of marker appearance that subjects patients to an additional imaging dose 1,5 datasets 12 , restricting broader applications. This project developed a deep-learning-based algorithm to segment gold fiducial markers implanted near the liver tumour, using multi-centre, multi-platform data to demonstrate feasibility for real-time clinical implementation. Material/Methods: A compact Convolutional Neural Network (CNN) was trained on 314,625 kilovoltage (kV) X-ray images from twelve TROG 17.03 LARK trial patients, spanning five centres and two platforms and three respiratory motion-management techniques. A separate dataset of 505,163 images from 16 patients within three centres was used for validation. Evaluation metrics included sensitivity, specificity, and the area under the precision-recall curve (AUC). For end-to framework, 5,182 ground truth marker positions were manually segmented every 10 degrees of gantry rotation from 16 patients. Feasibility for clinical implementation was defined as markers identified within 2 mm of ground truth position on >95% of end testing through a simulated real-time Kilovoltage Intrafraction Monitoring (KIM) 13 segmentation approaches are either limited to datasets with small marker movements 9-11 or single centre

images and processing time less than 500ms 14

per image.

Results: The marker position was segmented by the CNN within 2 mm in 95.36% of frames on the X axis, and 97.65% within 2 mm on the Y axis (Figure 1) when assessed across multiple centres in the KIM framework. Sensitivity reached 98.63%, specificity was 99.94%, and the AUC was 0.9965. The processing time of 140+/-10ms (mean +/- SD) using an

NVIDIA GeForce RTX 3070 GPU fulfilled real-time requirements of sub-500ms latency 13

on multi-platform

data.