Click here to

Session: [Return to Session]

Real-Time Tracking Diaphragm Motion On During-Treatment KV Cone Beam Projection Images Using ResNet50

J Liang*, Q Liu, E Porter, D Yan, Beaumont Health System, Royal Oak, MI

Presentations

MO-H345-IePD-F5-4 (Monday, 7/11/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: Diaphragm manifested on Cone Beam projection image has been used as surrogate to monitor lung tumor position. We propose a deep learning method to detect diaphragm apex position of the ipsilateral lung in real-time on KV projections acquired during Lung SBRT delivery.

Methods: 7 hypo-lung SBRT patients with right lower lobe tumor were selected in this retrospective study. The imaging projection angle ranges from about -270 to 112 degrees. The right lung diaphragm apex was manually labeled as ground-truth on KV projections acquired during-treatment sessions with 899±167 projections per session. The backbone of our machine learning model is 50 layers convolution neural network (ResNet50). Four dense layers were added in the top layers. A weighted mean square function was used as the loss function. The model was trained on 5 patients’ data with image augmentation (random shift in two dimensions) to increase 4 times of the input data. The model is then applied to the 4 during-treatment sessions of the other 2 patients. The ground-truth and the predicted output of the model from the during-treatment projections were compared. The motion in Superior-Inferior direction was evaluated in this study.

Results: Absolute error for all the test projections was 1.98±2.00 mm, max error of 4.9mm for 95% projections with the largest error at the lateral projections. For the 4 individual sessions, the absolute error was 1.88±1.71, 2.07±1.72, 2.20±2.58 and 1.85±1.60mm respectively. There was no significant difference of error distribution across sessions. Predictions were generated in 18 milliseconds per projection when using one NVIDIA Quadro P6000 GPU card, well below 182 milliseconds projection image interval in Elekta XVI system.

Conclusion: The ResNet50 model was able to track diaphragm apex position, with the uncertainty < 5mm for 95% predictions, at ipsilateral lung on all projection angles.

Keywords

Image Guidance, Image-guided Therapy, Target Localization

Taxonomy

IM/TH- Image Registration Techniques: Machine Learning

Contact Email

Share: