3D UXFormer Fuses UNET and Transformer for Breast MRI Tissue Segmentation

A new dual-branch 3D UXFormer model combining 3D UNET and Vision Transformer outperformed existing algorithms for multi-tissue segmentation on breast MRI, achieving higher Dice and IoU scores.

• The model fuses local features from 3D UNET with global features from a Vision Transformer (ViT) via a dual-branch structure.
• The ViT branch was pre-trained using a Masked Autoencoder (MAE) to address data limitations.
• In experiments, 3D UXFormer achieved superior Dice and Intersection over Union (IoU) scores across different breast MRI datasets compared with state-of-the-art methods.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.