Purpose: To realize markerless pancreatic tumor tracking by using a deep learning model to detect the contour of the clinical target volume (CTV) on the orthogonal X-ray images acquired by the on-board imagers of the gimbaled linac system.
Methods: The input of the model was either the digitally reconstructed radiograph (DRR) image or the X-ray image acquired by the on-board imagers. The output was the image with the contour of the CTV. After the four-dimensional computer tomography was acquired, a data augmentation procedure contains interpolation between respiratory phases and tilt of gantry and ring angle was conducted to generated sufficient images to train the deep learning model. For one patient, a total of 2500 DRR images were generated for one X-ray tube angle. Two-thousand images were used to fine-tune the pre-trained model and the remaining images were used to evaluate the accuracy of the model. The contour of CTV was extracted from CTV-only DRR and then set as the ground truth of contour prediction. The dice similarity coefficient (DSC) and the centroid difference of contour at the isocenter level were used to evaluate the accuracy of prediction.
Results: We initially performed markerless pancreatic tumor contour prediction on the DRR images of one patient from orthogonal X-ray tube angles, 45° and 135°. The mean DSC between prediction and ground truth were 0.91 and 0.82 for 45° and 135°, respectively. The mean centroid differences were 2.10 mm and 5.25 mm, respectively. Running on a single NVIDIA GV100 graphics processing unit, the mean calculation time on an image was 172 ms.
Conclusion: This work demonstrates the success of the current workflow and adaption of the deep learning model on markerless tumor tracking for pancreatic cancer. The prediction accuracy and calculation time also indicate a bright future in clinical practice.