基于生成对抗网络的海马子区图像分割

Attribute to the small structures and the morphological complexity of the hippocampal subfields, it is hard to obtain desirable segmentation results with the traditional segmentation methods. Therefore, we introduce generative adversarial networks into image segmentation of hippocampal subfields. The introduced method can achieve the pixel-level segmentation of brain MR images. The generative model and the adversarial model are trained alternately. The approach was tested based on the brain MRI images of 32 volunteers from the CIND Center in San Francisco, USA. It was compared quantitatively and qualitatively with methods based on the sparse representation and dictionary learning and CNN. The results showed that the proposed method, which achieved a significant improvement in the segmentation accuracy of the hippocampal subfields, outperforms the existing methods based on the dictionary learning and sparse representation and CNN. The results reveal that the introduced method can effectively improve the segmentation accuracy of hippocampal subfields in the brain MRI images, which will provide the basis for the clinical diagnosis and treatment of neurodegenerative diseases.