Excitation-based fully connected network for precise NIR-II fluorescence molecular tomography

Area of Science:

• Biomedical Imaging
• Medical Physics
• Machine Learning in Medicine

Background:

• Fluorescence molecular tomography (FMT) enables 3D visualization of fluorescence biomarkers but is limited by simplified models and complex inverse problems.
• Tissue scattering and noise interference in conventional FMT hinder precise reconstruction of biomarker distribution.
• Developing advanced imaging techniques is crucial for accurate in vivo molecular imaging.

Purpose of the Study:

• To develop a novel NIR-II fluorescence molecular tomography (FMT) reconstruction strategy.
• To improve the precision and accuracy of 3D fluorescence biomarker distribution.
• To leverage machine learning for enhanced biomedical imaging applications.

Main Methods:

• Utilized second near-infrared (NIR-II) fluorescence imaging to minimize tissue scattering and noise.
• Developed an excitation-based fully connected network to model the inverse problem of NIR-II photon propagation.
• Integrated an excitation block for focused attention on light source-related neurons and added barycenter error to the loss function for improved localization accuracy.

Main Results:

• The proposed NIR-II FMT strategy demonstrated superior performance compared to baseline methods in numerical simulations.
• Validation through in vivo experiments confirmed the enhanced accuracy and precision of the novel reconstruction approach.
• The deep learning model effectively reconstructed the 3D distribution of the light source.

Conclusions:

• The novel NIR-II FMT reconstruction strategy significantly improves the accuracy of 3D fluorescence biomarker distribution.
• The developed deep learning model offers a robust solution for the inverse problem in FMT.
• This approach is expected to advance the application of machine learning in biomedical research and diagnostics.