Pyramid Convolutional Recurrent Network for Serial Medical Image Registration With Adaptive Motion Regularizations

Abstract

Objective: Serial medical image registration plays an important role in radiation therapy treatment planning. However, current deep learning-based deformable registration models suffer from excessive resource consumption and suboptimal precision issues. Moreover, the global regularization term may result in unrealistic deformations due to displacement field noise and intertissue sliding motion omission. Methods: This article proposes a patch-based pyramid convolutional recurrent neural network (pyramid CRNet) for serial medical image registration. Patch-wise training is employed to alleviate resource constraints. Incorporating spatiotemporal features across multiple scales is beneficial for focusing on more details to improve accuracy. Moreover, two motion adaptive techniques are introduced to provide anatomically plausible displacement fields. The first uses a guided filter to reduce noise and maintain motion continuity within organs. The second involves a pixel-wise weight regularization term within the loss function to provide a tailored solution for distinctive tissue characteristics, especially for sliding motion at organ boundaries. Results: Experiments were conducted on lung 4DCT images and cardiac cine MR images. Quantitative and qualitative results have demonstrated that our method can align anatomical structures across multiple images in a physiologically sensible manner. Conclusion: The significance of this work lies in its potential to address pressing challenges in clinical applications, and further investigations could be extended to explore different modalities and dimensions.