Multimodality Artificial Intelligence for Involved-Site Radiation Therapy: Clinical Target Volume Delineation in
Xin Tie1, Sarah A Milgrom2, Andrea C Lo3
1Department of Radiology, School of Medicine and Public Health, University of Wisconsin, Madison, Wisconsin.
Purpose:
Clinical target volume (CTV) delineation for involved-site radiation therapy (ISRT) in Hodgkin lymphoma (HL) is time-consuming because of the need to analyze multi-timepoint positron emission tomography (PET)/computed tomography (CT) scans coregistered to the planning CT. Our goal was to develop automated CTV segmentation algorithms that integrated multimodality imaging to facilitate ISRT planning.
Methods And Materials:
This study included planning CT, baseline PET/CT (PET1), and interim PET/CT (PET2) scans from 288 pediatric patients with high-risk HL 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 3 deep learning (DL) architectures (SegResNet, ResUNet, and SwinUNETR) and evaluated the impact of incorporating PET1 and PET2 images along with the planning CT. Performance was assessed using the Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95). Interobserver variability was estimated by comparing original institutional CTVs with those newly delineated by 4 radiation oncologists on 10 cases. The quality of CTVs generated by the top-performing model was independently assessed by radiation oncologists in 40 other cases using a 5-point Likert scale and compared against the original institutional CTVs.
Results:
In the external cohort, a SwinUNETR model incorporating planning CT, PET1, and PET2 images achieved the highest performance, with a DSC of 0.72 and HD95 of 34.43 mm. All models incorporating PET/CT images were significantly better (P < .01) than planning CT-only models. Interobserver variability analysis yielded a DSC of 0.70 and HD95 of 30.14 mm. In clinical evaluation, DL-generated CTVs received a mean quality score of 3.38 out of 5, comparable to original physician-delineated CTVs (3.13; P = .13) CONCLUSIONS: The DL model was able to generate clinically useful CTVs with quality comparable to manually delineated CTVs, suggesting its potential to improve physician efficiency in ISRT planning.
