(Saturday, 4/4/2020) [Mountain Time (GMT-6)]
Purpose: We aimed to develop a novel deep learning based framework for noninvasive real time tumor motion tracking with the X-ray imaging to facilitate the stereotactic body radiation therapy (SBRT) treatment of lung cancer.
Methods: A transfer learning model from the pretrained U-Nets was implemented to identify the lung tumor target on the extremely fuzzy sequential X-ray projection images and further track the tumor motion during the SBRT treatment of lung cancer patients. Randomized partitioning was applied to the model training and cross validations besides data augmentation. Deep supervision was integrated into the improved U-Nets to boost the feature representation for target extraction and mathematical morphological operations were further applied to retrieve the probability maps for the final target segmentation. The training dataset include retrospectively selected 20 X-ray projection sets (each has ~600 images) acquired on the TrueBeam Clinac (Varian Medical Systems, Palo Alto, CA) for the cone beam CT reconstruction. The results from the proposed model were compared to those by manual delineation using sensitivity, specificity, Dice similarity coefficients (DSC), Jaccard index (JAC) and Hausdorff distance (HD). The tracking accuracy was calculated by subtracting the predicted position from the reference.
Results: The median DSC, sensitivity, specificity, and JAC were 0.93, 0.885, 0.948, 0.793, respectively, compared to the manual delineation. The corresponding average HD was found to be 3.15 ± 1.43 mm. The average tracking accuracy for five lung cancer patients was 1.73 ± 0.82 mm. The computation time for each projection was less than 30ms.
Conclusion: The proposed U-Nets based transfer learning algorithm has shown its capacity to reliably identify the tumor target on the low contrast X-ray projection images and could offer a markerless tumor tracking solution to the clinically challenging yet crucial problem encountered during radiation treatment of thoracic cancer.
Contact Email
