Deep background-mismodeling-learned reconstruction for high-accuracy fluorescence diffuse optical tomography.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Computational Science

Background:

• Accurate image reconstruction in diffuse optical tomography (DOT) is often hindered by background mismodeling.
• Fluorescence diffuse optical tomography (FDOT) specifically requires precise modeling of background signals for high-fidelity results.

Purpose of the Study:

• To present a novel deep learning framework for high-accuracy FDOT reconstruction.
• To address and automatically correct background mismodeling errors within the reconstruction process.

Main Methods:

• A learnable regularizer incorporating background mismodeling constraints was formulated.
• A physics-informed deep network was employed to implicitly learn the background mismodeling.
• A deep-unrolled FIST-Net was designed for L1-FDOT optimization, minimizing learning parameters.

Main Results:

• The deep background-mismodeling-learned reconstruction framework significantly improved FDOT accuracy.
• Implicit learning of background mismodeling was validated through experimental results.
• The framework demonstrated effectiveness in enhancing image reconstruction quality.

Conclusions:

• The proposed deep learning approach offers a robust solution for accurate FDOT.
• This framework provides a generalizable method for improving image modalities with unknown background modeling errors.
• The study validates the power of physics-informed deep learning in addressing complex inverse problems.