Self-supervised bi-directional mapping generative adversarial network for arbitrary-time longitudinal interpolation of missing data

Abstract

Early prediction can assist in diagnosis and slow the progression of brain diseases. As the disease progresses, patients with brain diseases experience cerebral atrophy, and existing brain disease prediction methods based on structural MRI utilize manually extracted morphological change features. Due to the frequent occurrence of missing data in longitudinal MRI sequences and the scarcity of densely annotated atrophy information in existing longitudinal MRI datasets, supervised learning for brain atrophy is challenging. This paper proposes an automated method for learning morphological changes in MRI over the course of a disease, named BM-GAN. It employs a self-supervised approach that jointly learns the brain’s non-rigid deformation over time during the interpolation process and guides the interpolation generator through a bidirectional mapping module to produce missing MRIs consistent with disease progression. BM-GAN generates complete MRI sequences for the ADNI and OASIS dataset, and experimental results show competitive performance on image quality metrics. Moreover, existing disease classification methods based on SVM/CNN/3DCNN have seen an improvement in precision by 6.21% to 16% for AD/NC classification and 7.34% to 21.25% for AD/MCI/NC classification after using synthetic data generated by BM-GAN. Visual results indicate that BM-GAN can generate MRIs consistent with the brain atrophy trend of Alzheimer’s disease, thereby facilitating the prediction of brain diseases.