Unsupervised Learning-Based CBCT-CT Deformable Image Registration for CBCT-Guided Abdominal Radiotherapy

Daily Cone beam CT (CBCT) imaging provides necessary anatomical information for accurate patient setup. Image quality of CBCT is usually far inferior to simulation CT scans. A workaround is to register the CT to the CBCT such that the contours and Hounsfield Unit (HU) values of the CT can be propagated to the CBCT. However, the inconsistent HU values across CT and CBCT make it less effective to use conventional image similarity measures. We aim to develop an unsupervised registration network to overcome this challenge in multimodal CT-CBCT image registration.We propose to integrate directional local structural similarity into an unsupervised learning framework to perform abdominal CT-CBCT image registration. Directional local structural similarity measures the image's self-similarity which reflects the underlying structural similarity regardless of the modality in use. The CBCT and CT images were separately processed to extract directional local structural similarity feature maps in different directions. We concatenated the directional local structural similarity feature maps and the original images as network input. Taking both the original images and their respective structural similarity feature maps as input allows the network to fully explore the potential correlations between CBCT and CT for accurate deformation vector field (DVF) prediction. Salient features learnt through previous iterations were highlighted by attention gates across layers to expedite the learning process. A 3D bicubic interpolation was used to up-sample and smooth the predicted DVF. We performed a leave-one-out cross validation with an image dataset of 45 patients to evaluate the proposed registration method. Normalized cross correlation (NCC) and target registration error (TRE) between CBCT and deformed CT were calculated to quantify the registration accuracy.Our results show that the alignment between the abdominal soft tissues has been greatly improved after registration for all patients. The mean and standard deviation of NCC and TRE were 0.97 (range 0.95-0.99) and 1.88 (range 1.03-2.67) mm. The proposed network allows for many datasets to be used as training datasets since ground truth DVF is not needed for the training process. The proposed network can predict a final DVF via a single forward prediction, which is faster than the conventional iterative registration algorithms.We have developed a novel unsupervised multimodal image registration method for CT-CBCT abdominal image registration, which does not need ground truth DVF for training, and demonstrated its feasibility. Taking both CBCT and CT images and their respective directional local structural similarity features as input, the proposed network performs direct DVF prediction to register the abdominal CT to CBCT images. This tool will be useful for future CBCT-guided radiotherapy of abdominal malignancies.