Deep learning enabled multiscale optical phase imaging for biomedical applications

Goswami, Neha

Permalink

https://hdl.handle.net/2142/120393 Copy

Description

Title

Deep learning enabled multiscale optical phase imaging for biomedical applications

Author(s)

Goswami, Neha

Issue Date

2023-04-25

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

Anastasio, Mark A.

Doctoral Committee Chair(s)

Anastasio, Mark A.

Committee Member(s)

• Cunningham, Brian T.
• Kong, Hyunjoon
• Boppart, Stephen A.

Department of Study

Bioengineering

Discipline

Bioengineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Quantitative phase imaging
• Deep learning

Abstract

The smaller the sample, the higher the effects of perturbation will be on the associated measurements. Biological samples are highly sensitive to any such perturbation that has the potential to alter the natural state of such samples. Fluorescence microscopy is a great tool for high-resolution visualization of biological samples. However, it is accompanied by risk of toxicity by stains, phototoxicity by excitation light and photobleaching by long-term exposures. For a sensitive characterization of samples, these effects can introduce unknown variables in the measurements, which is not desirable at low scales of observations where such perturbations can rival the true signal itself. Label-free quantitative phase imaging (QPI) is a promising method that can provide unperturbed structural as well as phase measurements. However, QPI suffers from a major drawback which is specificity. QPI alone cannot differentiate between the structures of interests in the sample. To impart specificity to QPI, computational techniques involving deep learning are emerging. One such technique developed in our lab is called phase imaging with computational specificity (PICS). PICS involves training deep-learning models on a pair of QPI and fluorescence images to learn the specificity. These trained models are then used to infer specific structures from the QPI images of the unlabeled samples. In this thesis, we present a multi-scale investigation of biological samples and report the development of associated methods for applications in virology (sample size in nanometers), mitochondria assessment in cancer biology (sample size in tens of micrometer) and embryology (sample size ~100-150 micrometers). Towards the end of this thesis, we also present an introductory technique for improving the space-bandwidth product of a QPI system. Our focus in virology associated QPI studies is divided into two directions: development of clinical pathogen detection tools and applying QPI to understand viruses and virus-host interactions. We demonstrate the potential of PICS in detecting and classifying four viruses: SARS-CoV-2, H1N1, HAdV and ZIKV. We then report the results of clinical testing of the developed method to detect SARS-CoV-2 in the breath sample of patients that achieved an accuracy of 62%. Next, we turn our attention to the biological aspects of QPI-based virology investigations. We present the results of our study involving monitoring and characterization of reactivation of HIV in JLat cells. By using QPI, we were able to characterize the changes in the infected cells in association with the reactivation of HIV. We observed that on-an-average, the cells with reactivated HIV have a higher population-based temporal mean diameter and dry mass as compared to the cells with latent HIV. In a different study we also report the detection and characterization of bacteriophage viruses through QPI and characterize the effects of bacteriophage infection of E. coli. We note that the infection mechanism is dependent on the microenvironment (media) of the cells. For the mitochondria in cancer cell research aspect of this thesis, we present a deep learning-based label-free mitochondria detection method. We demonstrate the generalizability of our method to different cell lines and provide four applications where our tool can be deployed for mitochondrial investigations. We demonstrate that our model can successfully characterize changes in the mitochondria upon treatment with anti-cancer drugs, mitochondria modulator reagents and during cell-cycle. We were also able to find differences in the dynamics of the mitochondria from the whole cell and found that mitochondria are less diffusive in motion as compared to the overall cell dynamics. For the embryology part of this thesis, we demonstrate a deep learning based nuclei detection model for the mouse embryo. We then extract and model new biomarkers from the nuclei that are efficient predictors of the embryo health. This information is used to develop a feature-based embryo health grading classifier that achieves an AUROC>0.88 when tested on out-of-distribution dataset. We also report the development of an image-based classifier that can achieve an AUROC>0.97 for the embryo health grading classification. The novelty of our method lies in the measurements that our system can perform. Our method is capable of providing 3D structural and compositional information as well as health-grade of the embryo. Motivated by the challenges faced during the data acquisition of various projects mentioned above involving an oil based high NA objective, we developed a computational method based on deep learning that can increase the space-bandwidth product of the imaging system. Although, the reported work is an exploratory study, we were able to demonstrate the increase in space bandwidth product at two magnifications 10x/0.3NA and 40x/0.95 NA. This proposed method can then retain the large field of view of a low NA objective with increased resolution. Applying this method to a 40x acquisition can lead to better resolution images along with the advantage of higher scan area and unperturbed long-time lapse imaging without the need of an oil objective.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120393

Copyright and License Information

Copyright 2023 Neha Goswami

Owning Collections