國立虎尾科技大學 |

Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium./
作者:	Jiang, Wenjun.
面頁冊數:	1 online resource (173 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
Contained By:	Dissertations Abstracts International85-11B.
標題:	Fluid mechanics. -
電子資源:	click for full text (PQDT)
ISBN:	9798382757346

Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium.
Jiang, Wenjun.

Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium. - 1 online resource (173 pages)

Source: Dissertations Abstracts International, Volume: 85-11, Section: B.

Thesis (Ph.D.)--New York University Tandon School of Engineering, 2024.

Includes bibliographical references

Mathematical models of radiative transfer are complex due to the integro-differential nature of the governing equations. Forward problems, where the radiation intensities are predicted for known domains and boundary conditions, are computationally complicated and expensive. Inverse problems are orders of magnitude more difficult, and implementation of traditional mathematical inversion techniques either fail to yield viable solutions or yield low fidelity reconstructions. In recent years, machine learning has shown significant potential in solving a variety of complex problems, including obtaining solutions where other methods fail as well as improving the accuracy of available results.This dissertation investigates the feasibility of using neural network models to obtain solutions of inverse problems in radiative transfer in an absorbing-scattering medium. As a representative implementation, a diffuse optical tomography (DOT) system that uses visible and near-infrared light is simulated to study the capability of neural networks to reconstruct images. The inverse problem of DOT image reconstruction represents a significant challenge in medical imaging and optical sciences due to its nonlinear, ill-posed, and underdetermined nature. Traditional DOT reconstructions suffer from low reliability and accuracy, and are computational inefficient.Several neural network models have been developed to solve the inverse problems in DOT. The location, size, optical properties (LSOP) model predicts the critical parameters of the target anomaly in the DOT system, such as the location, size, and optical properties, instead of reconstructing the entire image, provided the shape and number of anomalies are known. While the LSOP model has the benefit of a faster training process and improved accuracy compared to the traditional reconstruction method, it is challenging to adapt it to more generalized cases. A full pixel-by-pixel image reconstruction model is therefore developed that reconstructs the entire domain. To improve accuracy and fidelity of reconstruction a customized loss function is also developed for the fully connected neural network (FCNN). The customized loss function embeds the idea of peak signal-to-noise ratio (PSNR) and image correlation coefficient (ICC), both of which are measurements of image similarity. This ICC-PSNR hybrid model offers significant advantages compared to traditional reconstruction methods and neural networks with standardized loss functions like mean squared error (MSE). The proposed ICC-PSNR model not only highly improves reconstruction accuracy but is also robust in handling noisy detector data and capable of reconstructing target anomalies with irregular shapes.Further investigations are performed for more challenging situations that incorporate the most recent developments in neural networks. Physics-informed neural networks (PINN) and latent diffusion models (LDM) are used to improve the reconstruction results further. These are found to be especially significant for usually intractable cases such as those with very low number of detectors and detector-source pairs, as well as faint anomalies. The present study demonstrates the feasibility and strength of neural network models for high-fidelity image reconstruction of scattering absorbing mediums with inhomogeneities. In medical applications, these models have the potential of making DOT an accurate imaging modality that has remained elusive over the decades since DOT was proposed as a safer imaging alternative. It also provides a powerful tool for other applications such as in astrophysics and terrestrial imaging.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798382757346Subjects--Topical Terms:

555551
Fluid mechanics.
Subjects--Index Terms:

Inverse problemIndex Terms--Genre/Form:

554714
Electronic books.

Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium.
LDR:05079ntm a22003977 4500 001 1150322
005 20241028114746.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798382757346
035 $a (MiAaPQ)AAI30990962
035 $a AAI30990962
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Jiang, Wenjun. $3 1476784
245 1 0 $a Neural Networks for Inverse Problems of Radiative Transfer in Absorbing-Scattering Medium.
264 0 $c 2024
300 $a 1 online resource (173 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
500 $a Advisor: Kumar, Sunil.
502 $a Thesis (Ph.D.)--New York University Tandon School of Engineering, 2024.
504 $a Includes bibliographical references
520 $a Mathematical models of radiative transfer are complex due to the integro-differential nature of the governing equations. Forward problems, where the radiation intensities are predicted for known domains and boundary conditions, are computationally complicated and expensive. Inverse problems are orders of magnitude more difficult, and implementation of traditional mathematical inversion techniques either fail to yield viable solutions or yield low fidelity reconstructions. In recent years, machine learning has shown significant potential in solving a variety of complex problems, including obtaining solutions where other methods fail as well as improving the accuracy of available results.This dissertation investigates the feasibility of using neural network models to obtain solutions of inverse problems in radiative transfer in an absorbing-scattering medium. As a representative implementation, a diffuse optical tomography (DOT) system that uses visible and near-infrared light is simulated to study the capability of neural networks to reconstruct images. The inverse problem of DOT image reconstruction represents a significant challenge in medical imaging and optical sciences due to its nonlinear, ill-posed, and underdetermined nature. Traditional DOT reconstructions suffer from low reliability and accuracy, and are computational inefficient.Several neural network models have been developed to solve the inverse problems in DOT. The location, size, optical properties (LSOP) model predicts the critical parameters of the target anomaly in the DOT system, such as the location, size, and optical properties, instead of reconstructing the entire image, provided the shape and number of anomalies are known. While the LSOP model has the benefit of a faster training process and improved accuracy compared to the traditional reconstruction method, it is challenging to adapt it to more generalized cases. A full pixel-by-pixel image reconstruction model is therefore developed that reconstructs the entire domain. To improve accuracy and fidelity of reconstruction a customized loss function is also developed for the fully connected neural network (FCNN). The customized loss function embeds the idea of peak signal-to-noise ratio (PSNR) and image correlation coefficient (ICC), both of which are measurements of image similarity. This ICC-PSNR hybrid model offers significant advantages compared to traditional reconstruction methods and neural networks with standardized loss functions like mean squared error (MSE). The proposed ICC-PSNR model not only highly improves reconstruction accuracy but is also robust in handling noisy detector data and capable of reconstructing target anomalies with irregular shapes.Further investigations are performed for more challenging situations that incorporate the most recent developments in neural networks. Physics-informed neural networks (PINN) and latent diffusion models (LDM) are used to improve the reconstruction results further. These are found to be especially significant for usually intractable cases such as those with very low number of detectors and detector-source pairs, as well as faint anomalies. The present study demonstrates the feasibility and strength of neural network models for high-fidelity image reconstruction of scattering absorbing mediums with inhomogeneities. In medical applications, these models have the potential of making DOT an accurate imaging modality that has remained elusive over the decades since DOT was proposed as a safer imaging alternative. It also provides a powerful tool for other applications such as in astrophysics and terrestrial imaging.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Fluid mechanics. $3 555551
650 4 $a Computer engineering. $3 569006
650 4 $a Mechanical engineering. $3 557493
653 $a Inverse problem
653 $a Machine learning
653 $a Neural network
653 $a Radiative transfer
653 $a Diffuse optical tomography
655 7 $a Electronic books. $2 local $3 554714
690 $a 0548
690 $a 0464
690 $a 0204
710 2 $a New York University Tandon School of Engineering. $b Mechanical & Aerospace Engineering. $3 1468400
710 2 $a ProQuest Information and Learning Co. $3 1178819
773 0 $t Dissertations Abstracts International $g 85-11B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30990962 $z click for full text (PQDT)