Monotone Accelerated Proximal Gradient Network For Bioluminescence Tomography Reconstruction.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Computational Biology

Background:

• Bioluminescence tomography (BLT) enables 3D visualization of tumor cells but faces challenges from photon scattering and ill-posed inverse problems.
• Deep learning offers potential for optical tomography reconstruction, yet parameter selection and interpretability remain issues.
• BLT data exhibits non-Euclidean spatial relationships, complicating reconstruction.

Purpose of the Study:

• To develop a novel deep learning network for improved bioluminescence tomography reconstruction.
• To enhance the stability and interpretability of BLT reconstruction methods.
• To overcome limitations of existing regularization and deep learning approaches in BLT.

Main Methods:

• Proposed a novel Monotone accelerated proximal gradient network (MAPG-net).
• Combined advantages of regularization methods and Graph Attention (GAT) for BLT reconstruction.
• Inherited solution constraints from regularization frameworks to improve network stability and interpretability.

Main Results:

• The MAPG-net demonstrated excellent performance in numerical experiments.
• The network effectively addressed the ill-posed nature of BLT reconstruction.
• Improved stability and interpretability were achieved compared to traditional methods.

Conclusions:

• MAPG-net offers a promising approach for robust and interpretable bioluminescence tomography reconstruction.
• The integration of regularization and GAT provides significant advantages for 3D tumor visualization.
• This method has the potential to advance noninvasive cancer research and diagnostics.