Purpose: Current image-guided radiation therapy (IGRT) often relies on cone-beam computed tomography (CBCT). The difference between CBCT and kilovoltage computed tomography (kVCT) images makes it impossible to effectively evaluate the dose and adjust the treatment plan based on CBCT image. Current registration methods optimize an objective function independently for each pair of images, which can be time-consuming for extensive data. We are developing a novel unsupervised registration algorithm to generate deformation fields for dose assessment and plan adjustment without the need for manual re-scanning and re-planning.
Methods: We developed a registration network based on a generative adversarial network (GAN) with an attention gate(AG) to generate deformation fields aligned kVCT to CBCT images. Data from 37 liver cancer patients were collected, 30 of which were used for training and 7 for testing. For contours of left and right kidneys on patients in the test set, we compared different registration network structures with the evaluation metrics, including dice similarity coefficient (DSC) and deviation of each slicer’s centroid (DC) to try to find the most suitable network structure.
Results: Compared to rigid alignment, the mean DSC of each slicer of left and right kidneys changed from 0.862±0.031, 0.859±0.063 to 0.866±0.026, 0.863±0.053, and the mean DC of each slicer from 11.03±12.26, 6.92±4.90 to 3.37±3.65, 5.14±4.06 mm. Region of interest (ROI) coincidence rate was improved, and centroid distance was reduced.
Conclusion: With a deformation field generated by a small un-annotated data set for training, the aligned ROI of the kVCT images achieves similar to the results of CBCT images. It solves the problem of outlining and dose calculation in adaptive radiation therapy (ART). The registration technique fully meets the clinical requirements and is suitable for ART.