Application of neural networks to the reconstruction of diffuse optical tomography neuroimages

Granstedt, Jason Louis

Permalink

https://hdl.handle.net/2142/124655 Copy

Description

Title

Application of neural networks to the reconstruction of diffuse optical tomography neuroimages

Author(s)

Granstedt, Jason Louis

Issue Date

2024-04-15

Director of Research (if dissertation) or Advisor (if thesis)

Anastasio, Mark A

Doctoral Committee Chair(s)

Anastasio, Mark A

Committee Member(s)

• Forsyth, David A
• Wang, Yuxiong
• Culver, Joseph P

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Neural networks
• Deep learning
• Image reconstruction
• Diffuse optical tomography
• Neuroimaging

Abstract

Neuroimaging has led to several advancements in understanding brain functionality and neurological disorders. Diffuse optical tomography (DOT) offers a safe and portable neuroimaging modality. However, the resolution of DOT lags behind functional magnetic resonance imaging (fMRI). One avenue for increasing this resolution is via more advanced reconstruction methods. Neural networks have recently achieved state-of-the-art performance for many reconstruction problems, but conventional methods require a set of paired images for training not currently available for DOT. This project proposes a method for improving the resolution of DOT images via neural network reconstruction techniques. A dataset of paired images for training networks is constructed through exploiting the correlation between DOT and fMRI neuroimages. A virtual imaging pipeline was constructed to simulate DOT measurements from fMRI images and applied to a dataset with retinotopic stimuli. An SVD and null space analysis were performed to characterize the DOT imaging system. Three state-of-the-art network architectures are modified to support the ill-conditioned low-rank DOT imaging operator and employed for the reconstruction problem. Hallucination maps and in-distribution Rayleigh tasks were developed to properly assess the quality of the network reconstructed images and compare them with conventional methods. The networks trained on simulated data were then adapted to in-vivo DOT data and compared to the current state-of-the-art reconstruction approach. Statistically significant decoding results on in-vivo DOT data were observed for several models for the considered DOT imaging systems, indicating that the developed methods are promising for future development of network-based reconstructions for DOT applications.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Jason Granstedt

Owning Collections