Multiscale feature fusion network for 3D head MRI image registration

• DOI: 10.1002/mp.16387
• Publication Date: 2023-03-27T08:26:04Z
• Authors: Shixin Yang; Haojiang Li; Shuchao Chen; Wenjie Huang; Demin Liu; Guangying Ruan; Qiangyang Huang; Qiong Gong; Lizhi Liu; Hongbo Chen
• Keywords: Image registration; Feature (linguistics); Rigid transformation; Hausdorff distance

Abstract Background Image registration technology has become an important medical image preprocessing step with the wide application of computer‐aided diagnosis in various analysis tasks. Purpose We propose a multiscale feature fusion based on deep learning to achieve accurate and head magnetic resonance imaging (MRI) solve problem that general methods cannot handle complex spatial information position MRI. Methods Our proposed network consists three sequentially trained modules. The first is affine module implements transformation; second realize non‐rigid transformation, deformable composed top‐down bottom‐up subnetworks parallel; third also realizes transformation two series. decomposes deformation field large displacement into multiple fields small by registration, which reduces difficulty registration. Moreover, MRI learned targeted manner, improves accuracy, connecting subnetworks. Results used 29 3D MRIs for training seven volumes testing calculated values evaluation metrics new algorithm register anterior posterior lateral pterygoid muscles. Dice similarity coefficient was 0.745 ± 0.021, Hausdorff distance 3.441 0.935 mm, Average surface 0.738 0.098 Standard deviation Jacobian matrix 0.425 0.043. achieved higher accuracy compared state‐of‐the‐art methods. Conclusions can end‐to‐end MRI, effectively cope characteristics rich details images provide reliable technical support diseases.