Author: Tyler J Bradshaw, Sharon M Castellino, Steve Y Cho, David Hodgson, Bradford S Hoppe, Kara M Kelly, Andrea Lo, Sarah Milgrom, Xin Tie 👨🔬
Affiliation: Department of Radiation Oncology, University of Toronto, Department of Radiology, University of Wisconsin, University of Colorado Anschutz, Department of Medical Physics, University of Wisconsin, Department of Radiation Oncology, Mayo Clinic, Department of Radiation Oncology, BC Cancer, Vancouver Center, Department of Radiology, University of Wisconsin - Madison, Roswell Park Comprehensive Cancer Center, Emory University School of Medicine 🌍
Purpose: Clinical target volume (CTV) delineation for involved-site radiation therapy (ISRT) in Hodgkin lymphoma (HL) is time-consuming due to the need to analyze multi-time-point PET/CT scans co-registered to the planning CT. Deep learning (DL) has the potential to streamline this task, but its feasibility remains unexplored. Our goal was to develop automated CTV segmentation algorithms that integrated longitudinal and multi-modality imaging to facilitate ISRT planning.
Methods: This study included planning CT, pre-chemotherapy PET/CT (PET1), and interim-therapy PET/CT (PET2) scans from 288 pediatric patients enrolled in the Children’s Oncology Group AHOD 1331 trial. Data from 58 patients across 24 institutions were held out for external testing, while the remaining 230 cases from 95 institutions were used for model development. We investigated three backbone architectures (SegResNet, ResUNet and SwinUNETR) and evaluated the impact of incorporating PET1 and PET2 images alongside the planning CT. Additionally, we explored multi-modality early fusion and late fusion techniques for integrating PET/CT images into a longitudinally-aware segmentation model. Performance was assessed using the 95th percentile Hausdorff distance (HD95), Dice score, and average symmetric surface distance (ASSD).
Results: On the external cohort, the SwinUNETR model with late fusion achieved the highest performance, with an HD95 of 34.43 mm, Dice score of 0.72, and ASSD of 7.37 mm. In comparison, the best planning CT-only model attained an HD95 of 58.94 mm, Dice score of 0.68, and ASSD of 10.71 mm. All models incorporating PET/CT images demonstrated significant improvements (P<0.05) over CT-only models. Subset analyses showed no significant performance differences across variables such as age, sex, initial staging, presence of B symptoms, treatment arms, or PET2 Deauville scores.
Conclusion: This study demonstrates the feasibility of automated CTV segmentation for ISRT, leveraging multi-time-point and multi-modality imaging. We explored a novel and challenging application of DL, showing its potential to streamline ISRT planning.