Unsupervised generative AI for enhancing brain tumor segmentation in multi-center, incomplete real-world data scenarios

The clinical application of deep learning (DL)-based brain tumor segmentation remains limited by missing MRI sequences and cross-center data inconsistencies. Existing supervised generative models can synthesize missing sequences but rely on paired, fully sampled training data, which are often unavailable in routine practice. This study aims to assess the use of an unsupervised generative model to complete missing sequences and eliminate cross-center data inconsistency, and to verify whether using the generated images can enhance brain tumor segmentation. We retrospectively evaluated 921 glioblastoma (GBM) patients from BraTS, UCSF-PDGM, and our institutional datasets, together with 1000 meningioma cases from BraTS-MEN cohort. We developed an unsupervised multi-center multi-sequence generative adversarial transformer (UMMGAT) to generate MRI sequences from incomplete datasets. Key features of UMMGAT include a sequence encoder that disentangles and encodes modality-specific characteristics, and a lesion-aware module (LAM) that enhances the generation of tumor regions, all trained via multi-task learning for generating multi-modal images. Validation on GBM and meningioma segmentation task demonstrated that generated MRI sequences significantly improved segmentation performance across various missing-sequence scenarios. The enhancement in segmentation performance when T1ce was missing is an improvement that previous methods have not achieved. Further validation on an external local dataset confirmed that UMMGAT effectively adapts to cross-center data variations. With minimal training data requirements and the ability to generate multi-sequence MRI across multiple centers, UMMGAT provides a practical solution for handling incomplete and heterogeneous MRI data, facilitating more consistent and accurate brain tumor segmentation in diverse clinical contexts.