Fluorescence molecular tomography(FMT)can non-invasively monitor glioblastomas in small animals.Both handcrafted prior regularization and deep learning algorithms have made remarkable achievements in this field.But ha...Fluorescence molecular tomography(FMT)can non-invasively monitor glioblastomas in small animals.Both handcrafted prior regularization and deep learning algorithms have made remarkable achievements in this field.But handcrafted priors often cannot deal well with different kinds of tumors.Also,some deep learning methods still rely on handcrafted priors.In this paper,we introduce a Chebyshev non-explicit prior regularizer network(CNPRN).It replaces the handcrafted prior with a non-explicit prior and combines it with an optimization-inspired network.The CNPRN has two main parts:First,because of the long-range spatial correlation of light transmission in the finite element mesh,we create a non-explicit prior regularizer using high-order Chebyshev graph convolution.We also add inter-stage information pathways to combine useful data from the reconstructed outputs of each phase's regularizer;Second,to solve the problem of heavy computation in iterative optimization and make the network more flexible,we introduce a dynamic gradient descent module.This module allows parameters to be adjusted adaptively.As a deep unrolling method,CNPRN naturally obtains the solution constraints of the half-quadratic splitting method.This improves the network's generalizability and stability.Both simulations and in vivo experiments indicate that CNPRN has superior reconstruction performance.展开更多
基金supported by the National Natural Science Foundation of China(Nos.12271434,62271394,and 62201459)the Natural Science Basic Research Plan in Shaanxi Province of China(No.2023-JC-JQ-57)the Scientific and Technology New Star in Shaanxi Province of China(No.2023KJXX-035)。
文摘Fluorescence molecular tomography(FMT)can non-invasively monitor glioblastomas in small animals.Both handcrafted prior regularization and deep learning algorithms have made remarkable achievements in this field.But handcrafted priors often cannot deal well with different kinds of tumors.Also,some deep learning methods still rely on handcrafted priors.In this paper,we introduce a Chebyshev non-explicit prior regularizer network(CNPRN).It replaces the handcrafted prior with a non-explicit prior and combines it with an optimization-inspired network.The CNPRN has two main parts:First,because of the long-range spatial correlation of light transmission in the finite element mesh,we create a non-explicit prior regularizer using high-order Chebyshev graph convolution.We also add inter-stage information pathways to combine useful data from the reconstructed outputs of each phase's regularizer;Second,to solve the problem of heavy computation in iterative optimization and make the network more flexible,we introduce a dynamic gradient descent module.This module allows parameters to be adjusted adaptively.As a deep unrolling method,CNPRN naturally obtains the solution constraints of the half-quadratic splitting method.This improves the network's generalizability and stability.Both simulations and in vivo experiments indicate that CNPRN has superior reconstruction performance.
