Sparse‐graph manifold learning method for bioluminescence tomography

Abstract

In preclinical researches, bioluminescence tomography (BLT) has widely been used for tumor imaging and monitoring, imaged-guided tumor therapy, and so forth. For these biological applications, both tumor spatial location and morphology analysis are the leading problems needed to be taken into account. However, most existing BLT reconstruction methods were proposed for some specific applications with a focus on sparse representation or morphology recovery, respectively. How to design a versatile algorithm that can simultaneously deal with both aspects remains an impending problem. In this study, a Sparse-Graph Manifold Learning (SGML) method was proposed to balance the source sparseness and morphology, by integrating non-convex sparsity constraint and dynamic Laplacian graph model. Furthermore, based on the nonconvex optimization theory and some iterative approximation, we proposed a novel iteratively reweighted soft thresholding algorithm (IRSTA) to solve the SGML model. Numerical simulations and in vivo experiments result demonstrated that the proposed SGML method performed much superior to the comparative methods in spatial localization and tumor morphology recovery for various source settings. It is believed that the SGML method can be applied to the related optical tomography and facilitate the development of optical molecular tomography. K E Y W O R D Sbiology and medicine, bioluminescence tomography, image reconstruction techniques, inverse problems

Regularized reconstruction based on joint smoothly clipped absolute deviation regularization and graph manifold learning for fluorescence molecular tomography

“…Therefore, employing SCAD regularization in FMT reconstruction yields sparser and more accurate solutions compared to using L0 or L1 regularization alone, resulting in improved FMT reconstruction accuracy. However, it is important to note that increasing sparsity in reconstruction results may lead to potential loss of morphological details (Guo et al 2020). Inspired by graph-based manifold learning and sparse representation theory, we propose a novel approach called SCAD-GML for regularized FMT reconstruction.…”

“…L 1 -norm regularization results in over-sparseness in reconstruction. To overcome these problems, some new methods based on joint regularization, such as sparse-graph manifold learning (SGML) (Guo et al 2020) and L 1 −L 2 norm regulation via difference of convex algorithm (L 1 −L 2 via DCA) algorithm (Zhang et al 2016) have been proposed.…”

Based on model-driven fast iterative shrinkage thresholding network for bioluminescence tomography reconstruction

“…[11][12][13][14] Although these traditional model-driven methods can achieve a high-quality reconstruction result and enjoy the advantage of interpretability, due to the iterative nature of the solutions and the handcrafted characteristics, they suffer from high computational complexity. 15,16 Fueled by the rise of deep learning, several data-driven deep neural methods have been recently proposed for BLT reconstruction by direct learning the inverse mapping from the body surface information to the internal bio-distribution. 17,18 Compared to the model-driven methods, the data-driven deep neural methods dramatically reduce time complexity due to their non-iterative nature of the solutions and achieve impressive reconstruction performance.…”

Bioluminescence Tomography Based on One-Dimensional Convolutional Neural Networks

“…Traditionally, based on the light propagation model in biological tissues, the inversion algorithm is used to recover the three-dimensional (3D) distribution of the internal bioluminescent sources that enables quantitatively monitoring the pathological and physiological changes of the biological entities (4). In the past decade, BLT has been widely applied in preclinical studies such as early detection of tumors, monitoring tumor growth, and metastatic spreading (5)(6)(7)(8).…”