Deformable image registration with multi-level self-cross transformer architecture for the combined radiotherapy treatment of cervical cancer patients

The combined radiotherapy treatment is critical for the cervical cancer. This study aims to present an accurate deformable image registration method, which is critical to estimate the combined dose of External Beam Radiotherapy (EBRT) and Intracavitary Brachytherapy (ICBT). We introduce an enhanced Transformer based deformable image registration method to register the CT scans of EBRT and ICBT, which is characterized by a novel multi-level self-cross Transformer architecture based registration network (SCT-Net). In order to quantitatively evaluate the registration performance, the average dice similarity coefficient (DSC) and Hausdorff distance (HD) of the rectum and bladder between EBRT and ICBT after deformable image registration are computed with 80 cervical cancer patients. Without the registration between CT scans of EBRT and the last ICBT, the mean DSC values for bladder and rectum are 0.65 and 0.43, and the HD95 values for bladder and rectum are 6.1 and 6.6, respectively. After the registration, the mean DSC values of SCT-Net, XMorpher, VoxelMorph for bladder are 0.76, 0.74 and 0.70, respectively. Meanwhile, the mean DSC values of SCT-Net, XMorpher, VoxelMorph for rectum are 0.54, 0.52 and 0.49, respectively. In addition, the mean HD95 values of SCT-Net, XMorpher, VoxelMorph for bladder are 4.4, 4.8 and 5.2, respectively. Moreover, the mean HD95 values of SCT-Net, XMorpher, VoxelMorph for rectum are 5.9, 6.1 and 6.3, respectively. Visualized registration results further demonstrate the superiority of the proposed SCT-Net. These experimental results effectively demonstrate the superiority of the proposed SCT-Net for the deformable registration between the EBRT and ICBT, which is critical for the accumulative dose calculation of combined radiotherapy treatment.