Click here to

Session: Motion Compensation in Adaptive Radiotherapy [Return to Session]

Development of Markerless Intra-Fraction Motion Prediction in Image-Guided Radiotherapy Based On Deep Learning

D Zhou1*, M Nakamura1, N Mukumoto2, Y Matsuo3, T Mizowaki3, (1) Kyoto University, Graduate School of Medicine, Department of Human Health Sciences, Kyoto, JP, (2) Osaka City University, Graduate School of Medicine, Department of Radiation Oncology, Osaka, JP, (3) Kyoto University, Graduate School of Medicine, Department of Radiation Oncology and Image-Applied Therapy, Kyoto, ,JP


WE-C930-IePD-F2-3 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: A deep learning-based markerless real-time tumor tracking method for lung was developed. The performance was evaluated with clinical dynamic tumor tracking records and images.

Methods: Ten patients with lung cancer treated with a gimbal-head radiotherapy system were included. The prescription dose was 50 Gy in 4 fractions. Seven to nine-port non-coplanar static beams were used, corresponding to 14-18 orthogonal kV X-ray tube angles. A data augmentation procedure was conducted to expand the 10 phases four-dimensional computed tomography (4DCT) data scale 225-fold. For each X-ray tube angle of a patient, 2250 digitally reconstructed radiograph (DRR) images with gross tumor volume (GTV) labeled were obtained and adapted to fine-tune the pre-trained target contour prediction model. After the fine-tunning, the patient and X-ray tube angle specific GTV contour prediction model was acquired. The performance of the model was tested with the orthogonal X-ray projection images during treatment. The predicted three-dimensional (3D) positions of the GTV were calculated based on the centroids of the contours in the orthogonal X-ray images. For a gimbal-head radiotherapy system, the 3D positions of GTV calculated with markers implemented during treatment were recorded and considered as ground truth position. The 3D deviation between the prediction positions and the ground truth were calculated to evaluate the performance of the model.

Results: The median GTV size and motion range were 7.42 (range, 1.18-25.74) cm³ and 22 (range, 11-28) mm. A total of 8383 3D positions was included in evaluation. The mean GTV contour prediction time was 55 ms per image. The overall median value of 3D deviation was 2.29 (interquartile range, 1.72-2.98) mm. The probability for 3D deviation smaller than 5 mm was 93.3%.

Conclusion: Regarding to the evaluation results and calculation efficiency, this work may provide a solution for markerless real-time tumor tracking for patients with lung cancer.


Image-guided Therapy, Lung, Respiration


TH- External Beam- Photons: Motion management - intrafraction

Contact Email