D2C-Morph: Brain regional segmentation based on unsupervised registration network with similarity analysis

Abstract

Brain regional segmentation is an image-processing approach widely used in brain image analyses. Deep learning models that perform segmentation alone play an important role in medical fields such as automatic diagnosis and prognosis prediction. This method is effective for rapid diagnosis and large-scale processing. However, spatial alignment between image data is required for accurate segmentation. We proposed D2C-Morph, which can jointly perform registration and segmentation through unsupervised learning. The proposed model emphasizes the features of each input through a dual-path network and is designed to use contrastive learning twice. In addition, we demonstrated that the performance of the decoder can be improved by using a correlation feature map that enhances the similarity of the feature maps between two inputs through a correlation layer. Our study demonstrates that the deformation field of the registration network can be utilized for segmentation to jointly perform image processing pipelines.