Non-rigid Registration of Biomedical Image for Radiotherapy Based on Adaptive Feature Density Flow
Xu Jinghua,Tao Mingzhe,Zhang Shuyou,Jiang Xue,Tan Jianrong
DOI: https://doi.org/10.1016/j.bspc.2021.102691
IF: 5.1
2021-01-01
Biomedical Signal Processing and Control
Abstract:This paper presents a non-rigid biomedical image registration method based on Adaptive feature density flow (AFDF). Considering the muhiplicative characteristics of metadata signal for instance Magnetic Resonance Imaging (MRI) with high-resolution and high-sensitivity, the Non-Local Means Filtering Adapted to Rician Noise (NLMr) is employed to remove Rician noise rather than Gaussian noise with higher Peak Signal to Noise Ratio (PSNR). The deformation information and local shape information are homogeneous, so a flexible dimension of Fast Local Self-Similarity (FLSS) feature descriptor can be constructed to obtain adaptive FLSS (AFLSS) feature descriptor with adaptive Angle and Radial intervals. On the basis of AFLSS, the Adaptive feature density flow (AFDF) is proposed by calculating the dense feature flow field via Belief Propagation algorithm and thereby further optimized by the Pyramid iterative method. Finally, the image registration is completed by using the bicubic interpolation algorithm according to the dense feature flow field. Aiming at radiotherapy positioning control through precise robust registration of biomedical images, experiments of human brain MR images are carried out. The comparative results with PSNR, Root Mean Square Error (RMSE) and normalize mutual information (NMI) show that AFDF performs better than the optical flow method and Scale Invariant Feature Transform (SIFT) flow method, especially for the biomedical scenario with unconscious creep deformation.