Accurate segmentation of multimodal MRI is essential for brain tumor diagnosis, yet remains hindered by domain shifts across imaging modalities and the scarcity of labeled data. We introduce GUDA-Net, a novel end-to-end framework that performs task-driven unsupervised cross-modality adaptation via segmentation-guided generative learning. Unlike traditional CycleGAN-based pipelines that treat modality translation and segmentation independently, GUDA-Net establishes a bidirectional learning loop, where the segmentation network supervises the synthesis process, enforcing anatomical and semantic consistency in the generated modality. Central to our design is a modality-adaptive generator trained with dual consistency constraints—capturing both low-level visual fidelity and high-level semantic alignment. This strategy enables structure-preserving translation and enhances cross-domain generalization. Evaluated on the BraTS 2018 dataset, GUDA-Net achieves a Dice score of 89.90% in T2-to-T1 segmentation, surpassing existing state-of-the-art methods and underscoring its potential for clinically deployable multimodal MRI analysis under missing modality conditions.

Person re-identification (Re-ID) means matching people across different camera views based on different locations. It is challengeing for person images where there are background clutter, pose variations, illumination changes, etc. Attention mechanisms have become attractive for person re-identification algorithms as they aim at strengthening discriminative features, which accord with the purpose of person re-id, i.e., learning discriminative features for diffirent pedestrains. Previous approches mostly learn attention by useing local convolutions which have limited receptive filds. In this paper, an Relation Aware Attention(RAA) module is proposed to address this issue. RAA infers attention maps along two dimensions, channel and spatial, then them are mutiplied to the feature as the output map. For each feature position, RAA harvests the pairwise relationship with others as its response. Furthermore, in order to grasp the structure information of global scope and the local apperance information, we stack the relations and the feature to lean the final attention with a convlutional model. We designe the experiment and compare it with the existing benchmark. The resulits show that our attention model can increas the abillity of feature representation.