353 - Patient-Specific Markerless Tracking of Pancreatic-GTV and Abdominal Organs-at-Risk Using Deep-Learning
Presenter(s)
A. Ahmed1, L. Madden1, M. Stewart1, A. Mylonas2, R. Brown3, G. Metz1, M. Shepherd1, C. Coronel1, L. Ambrose1, A. Turk1, M. Crispin1, A. Kneebone1, G. Hruby1, and J. Booth1,4; 1Northern Sydney Cancer Centre, Royal North Shore Hospital, Sydney, Australia, 2ACRF Image X Institute - Faculty of Medicine and Health, Sydney, Australia, 3St George Cancer Care Centre, Sydney, NSW, Australia, 4Institute of Medical Physics, School of Physics, University of Sydney, Sydney, NSW, Australia
Purpose/Objective(s): In pancreatic stereotactic body radiotherapy (SBRT), real-time motion management with magnetic resonance imaging has improved patient outcomes. For standard linacs, similar improvements could be realized through accurate real-time motion management using intra-fraction kV images. While fiducial-based target tracking can account for rigid motion, the motion of pancreatic target and nearby organs can be large, non-rigid and dosimetrically significant. Deep-learning (DL) approaches for target tracking in intra-fraction kV images have shown promise for prostate and lung tumors. Applying DL for direct tracking of pancreatic target and organs-at-risk (OARs) could enable accurate real-time motion management for standard linacs. In this study, we investigate a DL solution for markerless multi-organ tracking of pancreatic target and radiosensitive OARs such as duodenum.
Materials/Methods: Patient data from an ethics approved clinical trial for pancreatic SBRT was utilized to train and test a conditional generative adversarial network (cGAN) for the segmentation of pancreatic-GTV, pancreas head, full pancreas, duodenum, stomach, and small bowel in 2D intra-fraction kV images. The dataset comprised contoured planning CT, contoured pre-treatment cone-beam CT (pre-CBCT), and intra-fraction kV images from 25 patients, all images acquired during exhale breath-hold or gated at exhale. Labeled Digitally reconstructed radiographs (DRRs) were generated from planning CT of 19 patients (CT-DRRs) and CBCT of 6 patients (CBCT-DRRs). A population model was trained using 68,400 CT-DRRs. Patient specific models were created for 6 additional patients for 15 fractions by fine-tuning the population model with 720 CBCT-DRRs per fraction and were evaluated using triggered kV images. Predicted contours were compared to forward-projected pre-CBCT contours, with model performance assessed using the Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and the 95th percentile Hausdorff Distance (HD95).
Results: Table 1 reports the metrics for contour prediction. The overall mean DSC, MSD and HD95 were 0.90 ± 0.1, 1.1±1.9 mm, 3.4 ± 5.0 mm, respectively. The prediction time per contour was 26.1 ± 0.8 milliseconds.
Conclusion: A patient-specific, markerless multi-organ tracking approach was developed and evaluated on intra-fraction kV images, demonstrating feasibility for real-time pancreatic SBRT. Further validation on a multi-institutional scale with an expanded training and testing dataset is planned to enhance its robustness and clinical applicability.
Table 1: Performance evaluation of 2D intrafraction contour prediction with cGAN compared to CBCT contours (mean±1SD)| Structure | DSC | MSD (mm) | HD95 (mm) |
| Pancreas GTV | 0.93 ± 0.1 | 0.5 ± 1.0 | 1.3 ± 2.3 |
| Pancreas Head | 0.92 ± 0.1 | 0.8 ± 1.5 | 1.9 ± 3.2 |
| Pancreas | 0.90 ± 0.1 | 0.9 ± 1.3 | 3.3 ± 4.2 |
| Duodenum | 0.85 ± 0.2 | 1.5 ± 2.7 | 4.3 ± 5.8 |
| Small Bowel | 0.90 ± 0.1 | 1.7 ± 2.5 | 5.8 ± 6.9 |
| Stomach | 0.92 ± 0.1 | 1.1 ± 1.8 | 3.7 ± 4.4 |
| Overall Mean | 0.90 ± 0.1 | 1.1 ± 1.9 | 3.4 ± 5.0 |