Non-rigid magnetic resonance image registration for cervical cancer radiation therapy evaluation using hybrid features

Abstract

Background: A non-rigid cervical magnetic resonance (MR) image registration algorithm combining pixel intensity and local region gradient features was proposed in this study for cervical cancer radiation therapy (RT) evaluation. Materials and Methods: The method was based on the following main steps: (1) each patient was scanned 2 times. The first scan was before internal-beam RT, and second scan was about 3~4 weeks after internal-beam RT. (2) DoG salient points mixed with stochastically sampled points were used as keypoints, and pixel intensity and PCA-SIFT features around them were extracted to build a feature vector for each keypoint. (3) In non-rigid registration process, α-mutual information (α-MI) was used as similarity measure. The method was evaluated by 20 MR images acquired from 10 patients with biopsy-proven squamous cell carcinomas. Results: For cervical cancer, the deformation of tumor and organ between different MR image acquisitions was subject to several errors, including possible mechanical misalignment, respiratory and cardiac motion, involuntary and voluntary patient motion, bladder and bowel filling differences. To minimize these ambiguities, patients filled their bladder before scanning. The proposed hybrid features can effectively catch the bladder and bowel in MR images, and α-mutual information (α-MI) based non-rigid registration can effectively align two long time internal MR images. Conclusion: Non-rigid cervical MR image registration method using hybrid features on α-MI can effectively capture different tissues in cervical MR images. Accurately aligned MR images can assist cervical cancer RT evaluation process.