Multierror learning enhanced fluorescence diffusion tomography in reflection geometry

Accurate three-dimensional (3D) localization and sensing of intravital fluorescent probes are indispensable for elucidating neural circuit mechanisms and evaluating tumor dynamics. Fluorescence diffusion tomography in reflection geometry (rFDT) offers a powerful tool for circumventing object size limitations and enables volumetric functional imaging in deep tissues. However, due to its spatially nonuniform detection sensitivity in reflection geometry, imperfect photon transport models, and tissue heterogeneity, the photon diffusion paths are highly susceptible to unexpected perturbations. Here, we present multierror-learning-enhanced rFDT (MEL-rFDT) for precise 3D localization and sensing of intravital fluorescent probes at subcentimeter depths. By embedding physical priors derived from photon transport models and spatial attention into the deep network, MEL-rFDT adaptively compensates for various errors and depth-dependent detection sensitivity, thus enabling the high-fidelity reconstruction of intravital fluorescent probes trained with only hundreds of in silico samples. Ex vivo brain tumor and in vivo subcutaneous tumor imaging in mice demonstrated MEL-rFDT's unprecedented 3D localization and sensing accuracy, as well as its volumetric and functional generalization across samples, facilitating intraoperative pathology, dynamic imaging, and reliable healthcare decision-making.