Preliminary Feasibility Study of Imaging Registration Between Supine and Prone Breast CT in Breast Cancer Radiotherapy Using Residual Recursive Cascaded Networks

Abstract

Breast cancer is one of the most common malignancies in women. The prone position in Partial Breast Irradiation (PBI) can better protect the heart and lung during radiotherapy. Supine position is used for CT imaging during treatment planning. The posture change in these two different positions may cause large deformation of breast, which make breast registration become a great challenge. Existing registration approaches for supine and prone breast images mainly use biomechanical modeling and iterative deformable images registration methods. However, the ability of these methods to capture such large deformations is limited. To tackle these problems, we propose an end-to-end residual recursive cascade network (RRCN) for supine and prone breast images registration. Unlike traditional deep learning networks, an affine subnetwork and several deformable subnetworks are trained together, enabling cooperation between subnetworks. Moreover, by using residual network connection, we can accelerate registration speed and reduce radiation dose. Registration accuracy is evaluated by visualizing registered images and computing normalized cross correlation (NCC). The experiment results show that RRCN with an average NCC of 0.982 ± 0.010 outperform VoxelMorph with an average NCC of 0.769 ± 0.070 and Recursive Cascaded Networks (RCN) with an average NCC of 0.914 ± 0.063, demonstrating the superior performance of the proposed method for supine and prone breast image registration. Because accurate deformable registration for this large-scale deformation is of great importance to the success of breast cancer radiotherapy, RRCN method has a strong potential to be a promising tool for future clinical practice in breast cancer radiotherapy.

Keywords

• Breast cancer
• radiation therapy
• deformable image registration
• deep learning
• cascade