Local2Global: UNet with Hierarchical Attention Mechanisms for Improved MR Image Inpainting

The BraTS Inpainting Challenge aims to synthesize healthy brain tissue to replace tumor-affected regions, which are localized with masks, on 3D magnetic resonance imaging. This effort supports clinical and research applications by generating anatomically plausible reconstructions. In this paper, we propose Local2Global, a novel UNet-like architecture combining convolutional and multiple attention mechanisms to synthesize anatomically coherent healthy brain tissue. Our model comprises four encoder stages, where each stage is designed to progressively capture distinct levels of context from local patterns with the convolution stage to global context with full attention. This local-to-global strategy enables the network to leverage the advantages of each layer type while reducing the computational burden associated with processing 3D volumes. Experimental results on the BraTS Inpainting dataset demonstrate the effectiveness of our approach. The proposed model was evaluated during the challenge. Our model achieved an SSIM of 0.768 and a PSNR of 20.548 on the validation set, whereas an SSIM of 0.844 and a PSNR of 21.954 on the testing set. The code is available on https://github.com/ThEnded32/Local2Global.