I2I-Mamba: Multi-modal medical image synthesis via selective state space modeling

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Multi-modal medical image synthesis requires capturing complex tissue interactions across spatial distances.
• Convolutional Neural Networks (CNNs) excel at local details but struggle with long-range context.
• Transformers offer long-range context but face complexity-driven trade-offs.

Purpose of the Study:

• To introduce a novel deep learning framework, I2I-Mamba, for effective multi-modal medical image synthesis.
• To address the limitations of CNNs and transformers in capturing diverse contextual features.
• To leverage the state space modeling (SSM) framework for improved contextual understanding.

Main Methods:

• Developed I2I-Mamba, a hybrid residual architecture utilizing dual-domain Mamba (ddMamba) blocks.
• Employed complementary contextual modeling in image and Fourier domains.
• Incorporated novel SSM operators with a spiral-scan trajectory for enhanced angular isotropy and a channel-mixing layer.

Main Results:

• I2I-Mamba demonstrated superior performance in multi-modal medical image synthesis.
• The method effectively captures both local spatial precision and broad contextual information.
• Evaluations on multi-contrast MRI and MRI-CT protocols confirmed its advantages over state-of-the-art methods.

Conclusions:

• I2I-Mamba offers a powerful new approach for multi-modal medical image synthesis.
• The ddMamba blocks and spiral-scan trajectory effectively model complex contextual features.
• This framework advances the capabilities of deep learning in medical image generation.