Latent diffusion-based image reconstruction for near-infrared spectral tomography

Area of Science:

• Biomedical optics
• Medical imaging
• Computational imaging

Background:

• Near-infrared spectral tomography (NIRST) offers non-invasive tissue optical property assessment.
• Image reconstruction in NIRST is challenging due to light scattering, leading to ill-posed problems and noise sensitivity.

Purpose of the Study:

• To develop an advanced image reconstruction algorithm for NIRST.
• To improve the accuracy and robustness of reconstructing chromophore concentrations from NIRST data.

Main Methods:

• Developed Diff-NIRST, integrating a conditional latent diffusion model, an autoencoding module, and a signal encoder.
• Utilized simulations for algorithm validation, reconstructing phantoms with varying inclusions and contrast ratios.
• Applied the trained algorithm to clinical breast cancer patient data.

Main Results:

• Diff-NIRST significantly outperformed Tikhonov regularization and other deep learning methods in simulations.
• Achieved substantial improvements in peak signal-to-noise ratio (PSNR) and reductions in contrast ratio (CR) and size error (SE) for hemoglobin and water content.
• Successfully reconstructed clinical images from patient data using a model trained solely on simulations.

Conclusions:

• Diff-NIRST demonstrates superior performance in NIRST image reconstruction compared to existing methods.
• The algorithm shows strong potential for translation to clinical practice, enabling more accurate non-invasive tissue analysis.
• The success with simulation-trained models highlights the robustness and adaptability of the Diff-NIRST approach.