(Wednesday, 7/15/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 2
Purpose: The purpose of this study is to develop a deep learning-based method to generate real-time patient-specific 3D volumetric images from any single 2D kV portal image for real-time tumor tracking during lung stereotactic body radiation therapy (SBRT).
Methods: We propose an inference generative adversarial network (InferGAN) with perceptual supervision to learn the patient-specific transformation between any single 2D kV image and 3D volumetric CT. Our network consists of three subnetworks, i.e., encoding, transformation, and decoding modules. Image distance loss between the generated 3D-CT images and the ground truth images is used to supervise these subnetworks. A pre-trained segmentation network is introduced to represent the generated and the ground truth 3D-CT images in the pyramid feature space. Perceptual loss in feature space is integrated into loss function to force the network to yield a sharper contrast around lung organs and tumors. To evaluate our method, we conducted a retrospective study with 20 lung SBRT cases, who had undergone 4D-CT simulation for motion evaluation purposes. For each 3D-CT image set of a breathing phase, we generated 2D projections at 179 different gantry angles. Then 3D-CT images with their corresponding 2D projections from 9 out of 10 phases were used as training data, with those from the remaining phase used for testing.
Results: The average mean absolute error, peak signal-to-noise ratio, normalized cross-correlation are 52.9±3.5 HU, 24.9±2.1dB and 0.98±0.01. The average target localization error is less than 2mm. Patient-specific 3D-CT images can be generated within 1 second via our trained network.
Conclusion: We have developed a novel deep learning-based real-time volumetric imaging method for lung SBRT, and demonstrated its feasibility and accuracy. This real-time imaging method could provide a potential solution for real-time lung tumor tracking, which is the basis for accurate dose delivery and further improving treatment accuracy and precision.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
