NIR-II fluorescence molecular tomography based on transformer encoder architecture

Fluorescence Molecular Tomography (FMT) is a highly sensitive method for depicting the three-dimensional distribution of fluorescent targets within biological tissues. However, traditional FMT reconstruction can be affected by the severe photon scattering and the modeling errors of the photon propagation model, making it difficult to obtain accurate reconstruction results. Deep learning methods can directly learn the nonlinear mapping between the photon intensity on the surface and the internal light sources, providing a promising approach for addressing the inverse problem in FMT. Since the second near-infrared window (NIR-II, 1000-1700) fluorescence imaging has a higher tissue penetration depth and imaging contrast, this work constructs a deep learning model suitable for the fluorescence images in this spectrum. The model introduced in this study, akin to grasping the contextual nuances of word vectors in the domain of machine translation, endeavors to comprehend the process of photon transmission within biological tissues. Inspired by the transformer encoder, we proposed a network mainly consist of positional encoding, dual-head attention mechanism, residual connection, and fully connected layers. It has achieved significantly better results than comparative methods on the NIR-II dataset constructed by the Monte Carlo method.