Multi-channel MRI reconstruction using cascaded Swinμ transformers with overlapped attention

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep neural networks excel at artifact reduction in magnetic resonance imaging (MRI) reconstruction.
• Attention-based vision transformer models increasingly outperform convolutional models in MRI reconstruction tasks.

Purpose of the Study:

• Investigate transformer architectures for multi-channel cascaded MRI reconstruction.
• Explore overlapped attention versus hybrid attention in shifted-window (Swin) transformers.
• Assess the impact of transformer layer count on reconstruction performance.

Main Methods:

• Utilized cascades of small transformers for multi-channel undersampled MRI reconstruction.
• Introduced and compared overlapped attention with hybrid attention in Swin transformers.
• Evaluated performance on standard 3T and low-field 0.3T T1-weighted MRI images at various acceleration rates.

Main Results:

• Models employing overlapped attention achieved superior or comparable quantitative metrics to state-of-the-art convolutional methods.
• Overlapped attention models demonstrated more consistent performance across different acceleration rates than hybrid attention models.
• Transformer architectures with fewer layers proved as effective as those with more layers in cascaded reconstruction.

Conclusions:

• Demonstrated the feasibility and effectiveness of cascaded small transformers with overlapped attention for MRI reconstruction.
• Achieved significant results without relying on pre-training on large external datasets like ImageNet.
• Highlighted the potential of novel transformer attention mechanisms for advanced medical image reconstruction.