Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Evaluation of Deep Learning-Based Algorithms for Tracking of Implanted Fiducial Markers in Pancreatic Cancer Patients

A Ahmed1, A Mylonas2, M Gargett1, L Madden1*, D Chrystall1, R Brown1, A Briggs1, D Nguyen2, P Keall2, A Kneebone1, G Hruby1, J Booth1, (1) Northern Sydney Cancer Centre, St Leonards, NSW, AU, (2) University of Sydney, Sydney, NSW, AU, (3) University of Technology Sydney, Ultimo, NSW


PO-GePV-M-198 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Real-time target position verification during pancreas stereotactic ablative radiotherapy (SABR) is important to detect irregular tumor motion. Fast and accurate fiducial marker segmentation is a requirement of real-time marker-based verification. Deep learning (DL) segmentation techniques are ideal because they don’t require additional learning imaging or prior marker properties (e.g., shape, orientation). In this study, we evaluate two DL methods for marker tracking applied to pancreatic cancer patient data.

Methods: The DL frameworks evaluated were 1) a CNN with sliding window classification and 2) a pretrained you-only-look-once (YOLO) version-4 architecture. CBCT projections and intrafraction kV imaging collected during pancreas SABR treatments were used as training data. All patients had 1-4 implanted fiducial markers (EchoTip® Ultra, Cook Medical). The performance of each model was tested on unseen data. The ground truth was calculated from manual segmentation and triangulation of markers in orthogonal paired kV/MV images. The root-mean-square error (RMSE) and standard error of mean (analyzed for 20 fractions from 10 patients) were calculated for the centroid of the markers predicted by the models, relative to the ground truth.

Results: The mean RMSE of the CNN was 0.61 ± 0.12 mm, 0.28 ± 0.04 mm and 0.15 ± 0.03 mm in the left-right, superior-inferior, and anterior-posterior directions, respectively, while the YOLO RMSE was 0.72 ± 0.33 mm, 0.18 ± 0.04 mm and 0.14 ± 0.02 mm respectively. The detection times of the models per frame on a GPU (NVIDIA GeForce RTX 3090) were 56.0 and 22.9 milliseconds for CNN and YOLO, respectively.

Conclusion: Two DL approaches were implemented to classify and track implanted fiducial markers in pancreatic cancer patient data. The accuracy and precision of marker position prediction by the DL models from the ground truth was submillimeter, and detection time was fast enough to meet the requirements for online application.

Funding Support, Disclosures, and Conflict of Interest: Patient data used for this study was collected under an ethically approved trial (ID: NCT03505229).


Not Applicable / None Entered.


TH- External Beam- Photons: Motion management - intrafraction

Contact Email