A model-driven deep unfolding network for fluorescence molecular tomography reconstruction

Abstract

As a high-sensitivity and high-specificity imaging method, fluorescence molecular tomography (FMT) can quantitatively reconstruct the distribution of fluorescence sources inside the organism, and has great application prospects in tumor diagnosis, medicine development, and treatment evaluation. However, the reconstruction accuracy of traditional FMT is limited by the oversimplified forward model and the severe ill-posedness of the inverse problem. A physical model-driven iteratively unfolding network named ISTA-UNet is proposed in this paper. By combining the model-driven Iterative Shrinkage/Thresholding (IST) process and the UNet network model, the ISTA-Unet framework can take advantage of the denoising and detail recovery capabilities of deep neural networks on the basis of guaranteeing interpretability. In order to verify the effectiveness of the network, this paper analyzes the reconstruction of fluorescent targets with different positions, edge-to-edge distances, and fluorescence yield ratios. The results demonstrates that the FMT reconstruction based on ISTA-UNet has a significant improvement in spatial resolution and quantification compared with traditional methods, and has great potential in improving the quality of image reconstruction.