Joint-optimization deep network for biomedical images registration and segmentation based on the double-sampling data augmentation

Biomedical image registration and segmentation are essential in life science research, medical diagnosis, and clinical treatment. However, deep learning methods typically rely on large-scale, high-quality annotated data, which are costly to obtain. The existing data augmentation techniques mainly generate additional samples through linear deformation such as rotation and translation methods, yet they still struggle to effectively solve the demand for data brightness diversity and elastic structure deformation diversity of biomedical images. In this paper, we propose a double-sampling data augmentation strategy based on the joint model of registration and segmentation. Integrating the double sampling on the brightness and deformation fields enriches the brightness diversity and the data's elastic structural deformation diversity. Further, to suppress the misleading of incorrect information in the pseudo-labeled data on the joint model, an adversarial training method is used to train a pseudo-label discriminator module to find inaccurate segmentation predictions and suppress the propagation of incorrect information in the data cycle. Eventually, this paper proposes a joint model for registration and segmentation for a one-shot scenario. It achieves better segmentation and registration performance than the comparison algorithms on the Mindboggle-101 human brain MRI dataset, MouseBrain mouse brain fMOST dataset, and MM-WHS 2017 heart dataset, demonstrating the effectiveness of our data augmentation strategy.