Program Information
Robust Tumor Identification by Convolutional-Deconvolutional Neural Networks for Motion Tracking On Thoracic Cone Beam CT Projections
M Chao1*, Y Yuan1 , J Wei2 , Y Lo1 , (1) The Mount Sinai Medical Center, New York, NY, (2) City College of New York, New York, New York
Presentations
SU-K-605-13 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: 605
Purpose: To develop a novel strategy to reliably identify the lung tumor on thoracic cone beam CT (CBCT) projections with deep learning techniques for markerless tumor motion tracking in stereotactic body radiation therapy (SBRT) of lung cancer.
Methods: The noisy and blurry lung tumor on CBCT projections from SBRT patients was automatically identified with 19-layer convolutional-deconvolutional neural networks (CDNN) which were trained end to end with ~200K parameters. This CDNN consists of two steps: 1) convolutional step: using repeated convolution and pooling to maximally capture the contextual feature information; 2) deconvolutional step: image resolution is restored through two inverse paths of deconvolution and upsampling. The training dataset include 100 CBCT projection image sets (each has 544 views) acquired on the Varian’s iX Clinac to train the parameters. To enable the CDNN training process, an essential supervised learning procedure, manual delineation of the tumor target on the same projection angle for half of the projections in the image set was completed, whereas the other half were segmented through interpolation of existing delineations. The trained model was evaluated on an independent dataset with ten patients. CDNN based results were compared to those by manual delineation using Dice Coefficients (DC), in addition to comparison with conventional tumor tracking techniques with prior information on the digitally reconstructed radiography (DRR) from planning CT.
Results: The average DC of the calculation by our deep learning method was between 0.9 and 0.95 compared to the manual delineation. The corresponding discrepancy was found to be 1.3 ~ 2.5mm as opposed to the average error of 3.2 mm from the conventional DRR-based tumor tracking algorithms.
Conclusion: The new CDNN-based deep learning algorithm demonstrated the capacity to reliably identify the tumor target on noisy CBCT projections and outperformed the traditional tumor tracking techniques based on prior information.
Contact Email: