Multi-operator-based model-driven self-supervised learning for fluorescence diffusion tomography.

Area of Science:

• Biomedical Imaging
• Optical Tomography
• Machine Learning

Background:

• Supervised learning in fluorescence diffusion tomography (FDT) is hindered by the requirement for extensive labeled data.
• Developing advanced imaging techniques for biological tissues necessitates overcoming data limitations.

Purpose of the Study:

• To introduce a novel self-supervised learning framework for FDT that eliminates the need for labeled data.
• To enhance the practical applicability of deep learning in experimental FDT.

Main Methods:

• A multi-operator-based model-driven self-supervised learning (MMSL) approach was proposed.
• Two forward operators derived from the photon transport model were integrated as dual constraints into an unrolled network.
• Geometrically disjoint source-detector configurations were exploited.

Main Results:

• MMSL achieved reconstruction quality comparable to supervised methods in experimental FDT.
• The method demonstrated superior recovery of morphological features compared to existing techniques.
• Successful validation was performed on a custom-built line-illumination FDT system.

Conclusions:

• The proposed MMSL method effectively addresses the data scarcity issue in FDT.
• This advancement broadens the scope of deep learning applications in experimental optical tomography.
• MMSL offers a viable alternative for high-fidelity imaging in scenarios lacking labeled datasets.