Self-Supervised Segmentation for Rigid 2D/3D Medical Image Registration

Tianhao Zhang, Xianhui Liu, Weidong Zhao

Abstract: Semantic segmentation is frequently utilized as a fundamental algorithm module in complex tasks. In the field of medical image registration, applying semantic segmentation to preliminarily identify the Regions of Interest (ROI) required for registration can provide data with reduced noise, thereby facilitating the training and registration process for model-based coarse registration. However, due to the difficulty in acquiring data in the medical imaging field, there is a scarcity of 2D/3D data pairs required for registration, making it challenging to form specific datasets for semantic segmentation training. Inspired by the use of Digitally Reconstructed Radiograph (DRR) for simulated registration, this paper utilizes 3D data and its annotations to provide simulated datasets for segmentation training based on the DRR algorithm. Additionally, it integrates pixel contrastive loss into the MedNeXt segmentation framework to enhance performance. Furthermore, it employs a ProST-based registration framework to validate the use of segmented data for registration, resulting in improved registration accuracy.

External IDs:doi:10.1145/3706890.3707001