Author: Fumiaki Komatsu, Shunsuke Moriya, Ryosuke Nakamura, Takeji Sakae, Toshiyuki Terunuma, Tetsuya Tomita 👨🔬
Affiliation: Graduate School of Comprehensive Human Sciences, University of Tsukuba, Institute of Medicine, University of Tsukuba, Proton Medical Research Center, University of Tsukuba, Department of Radiology, University of Tsukuba Hospital 🌍
Purpose: To develop a deep learning (DL) model capable of accurately tracking lung tumors independent of beam angle variations.
Methods: A thoracic dynamic phantom simulating lung motion in the superior-inferior direction was used. A simulated tumor (10 mm diameter, -150 HU) was positioned in the right lung, with a metal marker placed 7 cm from the tumor center. A treatment plan of six fields was created using non-coplanar fields on the CT image of the stationary phantom. Each field was irradiated with 500 MU at 600 MU/min using a 6 MV Linac while the phantom moved in a 4-second cosine wave pattern. The first 100 MU were delivered in an open 10 cm × 10 cm field for marker localization, with the remaining 400 MU shaped to the tumor. Cine EPID images acquired during irradiation were used for tracking. A DL model based on the Attention U-Net (AU-Net) architecture was implemented. Virtual EPID images were generated using DRRs of each field from the planning CT and MLC location data, with data augmentation applied through random MLC shifts and the Random Overlay method. A total of 3,000 paired DRRs and projected CTV contours (500 per field) were used for AU-Net training. Cine EPID images were input into the trained model frame-by-frame, and predicted contours were compared to ground truth contours derived from marker data.
Results: The mean positional error of tumor contours across all fields was 0.42 ± 0.24 mm. Frames with errors within 2 mm and 1 mm accounted for 100% and 95.5%, respectively. Contour agreement yielded a Jaccard index of 0.93 ± 0.03 and a 95% Hausdorff distance of 0.69 ± 0.25 mm.
Conclusion: The AU-Net model accurately tracked tumors across all beam angles, including non-coplanar fields, without requiring angle-specific training, demonstrating strong potential for clinical application.