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

Abstract

Supervised learning's reliance on high-fidelity labeled data limits its application in fluorescence diffusion tomography (FDT). Here, we propose a multi-operator-based model-driven self-supervised learning (MMSL) for FDT to eliminate the need for labeled data. Our approach exploits geometrically disjoint source-detector configurations to derive two forward operators from the photon transport model while integrating the operators as dual constraints into an unrolled network architecture: one enforces output-space consistency, and the other directs network parameter optimization. Experimental results on our custom-built line-illumination FDT system demonstrate that MMSL achieves reconstruction quality comparable to supervised methods while exhibiting superior recovery of morphological features. This advancement significantly expands the practical utility of deep learning in experimental FDT scenarios lacking labeled data.