Hybrid-CMLP: Hybrid CNN-MLP Networks for Low-to-standard-dose PET Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiological Physics

Background:

• Positron Emission Tomography (PET) is vital for disease detection but involves radiation exposure.
• Current PET synthesis methods using CNNs have limited receptive fields.
• Transformers struggle with computational costs at full resolution, while MLPs lack localized feature extraction.

Purpose of the Study:

• To develop a novel hybrid CNN-MLP network (Hybrid-CMLP) for synthesizing standard-dose PET (sPET) images from low-dose PET (lPET) data.
• To address limitations of existing methods by integrating global and local feature extraction efficiently.
• To improve the quality of synthesized PET images while minimizing patient radiation exposure.

Main Methods:

• Proposed a Hybrid-CMLP architecture featuring a novel Hybrid-Syn block.
• The Hybrid-Syn block integrates axis-wise MLP branches for global dependencies and dual CNN branches for local features.
• Introduced an Adaptive Fusion Mechanism (AFM) for dynamic integration of global and local features based on spatial context.

Main Results:

• Hybrid-CMLP demonstrated superior performance over state-of-the-art methods in PET synthesis.
• Experiments on two benchmark datasets showed consistent qualitative and quantitative improvements.
• The proposed architecture effectively captures both global anatomical/functional correlations and fine-grained tissue textures.

Conclusions:

• The Hybrid-CMLP network offers an effective solution for high-quality PET synthesis with reduced radiation exposure.
• The integration of CNNs and MLPs, along with the AFM, enhances the ability to model complex dependencies in PET data.
• This approach advances low-dose PET imaging, paving the way for safer and more accurate diagnoses.