Source distribution correlation enabled self-attention residual network for effective reconstruction of optical molecular tomography

As a promising preclinical imaging technique, optical molecular tomography (OMT) shows great potential in early detection and diagnosis of tumor diseases. However, its widespread application has been hindered by the limitations of traditional reconstruction methods, specifically the accuracy of optical transmission models and the ill-posed nature of inverse reconstruction. The development of deep learning has offered novel solutions for OMT, enabling efficient reduction of the ill-posed nature in reconstruction. The existing deep learning approaches employ conventional neural networks and objective functions, which retains significant scope for enhancing the accuracy of image reconstruction. In this paper, we propose a source distribution correlation enabled self-attention residual network (DCeSR network) to address the need for accurate OMT reconstruction. The DCeSR network leverages a residual learning strategy and a self-attention mechanism to effectively integrate the deep and shallow features, subsequently extracting highly informative surface measurements to accurately predict the three-dimensional distribution of light sources within tissues. The efficacy of the DCeSR network was validated through training and testing with two distinct numerical simulated datasets, each encompassing both single and dual source configurations. Both qualitative and quantitative analyses demonstrate the superior performance of the DCeSR network in achieving accurate OMT reconstructions.