Generative adversarial DacFormer network for MRI brain tumor segmentation.

Authors

Zhang M,Sun Q,Han Y,Zhang M,Wang W,Zhang J

Affiliations (5)

  • School of Computer Science and Engineering, Dalian Minzu University, Dalian, 116600, China.
  • School of Computer Science and Technology, Dalian University of Technology, Dalian, 116024, China.
  • Montreal Neurological Institute, Quebec, 116024, Canada.
  • School of Computer Science and Technology, Shandong Technology and Business University, Yantai, 264005, China.
  • School of Computer Science and Engineering, Dalian Minzu University, Dalian, 116600, China. [email protected].

Abstract

Current brain tumor segmentation methods often utilize a U-Net architecture based on efficient convolutional neural networks. While effective, these architectures primarily model local dependencies, lacking the ability to capture global interactions like pure Transformer. However, using pure Transformer directly causes the network to lose local feature information. To address this limitation, we propose the Generative Adversarial Dilated Attention Convolutional Transformer(GDacFormer). GDacFormer enhances interactions between tumor regions while balancing global and local information through the integration of adversarial learning with an improved transformer module. Specifically, GDacFormer leverages a generative adversarial segmentation network to learn richer and more detailed features. It integrates a novel Transformer module, DacFormer, featuring multi-scale dilated attention and a next convolution block. This module, embedded within the generator, aggregates semantic multi-scale information, efficiently reduces the redundancy in the self-attention mechanism, and enhances local feature representations, thus refining the brain tumor segmentation results. GDacFormer achieves Dice values for whole tumor, core tumor, and enhancing tumor segmentation of 90.9%/90.8%/93.7%, 84.6%/85.7%/93.5%, and 77.9%/79.3%/86.3% on BraTS2019-2021 datasets. Extensive evaluations demonstrate the effectiveness and competitiveness of GDacFormer. The code for GDacFormer will be made publicly available at https://github.com/MuqinZ/GDacFormer.

Topics

Brain NeoplasmsMagnetic Resonance ImagingNeural Networks, ComputerImage Processing, Computer-AssistedJournal Article

Ready to Sharpen Your Edge?

Join hundreds of your peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.