F Li*, W Cai, X He, P Zhang, L Cervino, X Li, T Li, Memorial Sloan Kettering Cancer Center, New York, NY

• 
  F Li

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-152 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: To investigate the feasibility and accuracy of a deep-learning based spine tracking algorithm using on-treatment kV imaging.

Methods: Varian’s on-treatment IMR allows visual inspection of patient movement during treatment, it however does not provide quantitative information about the motion. In this study, we trained a deep convolutional neural network to quantify the translational movement indicated by an IMR image. The proposed neural network registers an IMR to the corresponding spine-only DRR which is a pre-processed reference image representing the initial spine position established during patient setup. The neural network modifies Resnet101 to allow the input of a pair of images and regression for the relative translation (∆X, ∆Y) between them. The training and testing images were collected from nine patients treated for C-, T-, and L-spines, among which 390 images from eight patients were used for training and 41 images from the last patient were used for testing. To generate the full training data, we augmented the image set by randomly translating the IMR images to simulate patient movement. Each translated IMR is paired with the corresponding spine-only DRR to form an input sample, and the applied artificial translation is the training label.

Results: The trained neural network were evaluated on the test dataset with simulated spine movements up to 5mm. The mean detection error was below 0.7mm. The performance was compared to mutual information (MI) based matching implemented by Matlab. We found that our model is significantly more accurate and robust at predicting spine movements larger than 2mm, in which case MI-based method were undermined by inherent soft tissue interference.

Conclusion: The proposed deep convolutional neural network is capable of detecting spine motions with sub-millimeter accuracy for movements up to 5mm. Our results suggest that the proposed deep-learning approach is feasible for spine tracking during a SBRT treatment.

Funding Support, Disclosures, and Conflict of Interest: The present study is partially supported by a research agreement with Varian Medical Systems.

ePosters

Keywords

Image-guided Therapy, Quality Assurance

Taxonomy

IM/TH- Image Registration Techniques: Machine Learning

Contact Email