Fluorescence microscopy image analysis of retinal neurons using deep learning
<p>An essential goal of neuroscience is to understand the brain by simultaneously identifying, measuring, and analyzing activity from individual cells within a neural population in live brain tissue. Analyzing fluorescence microscopy (FM) images in real-time with computational algorithms is es...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| _version_ | 1811139700185890816 |
|---|---|
| author | Cudic, M |
| author2 | Noble, A |
| author_facet | Noble, A Cudic, M |
| author_sort | Cudic, M |
| collection | OXFORD |
| description | <p>An essential goal of neuroscience is to understand the brain by simultaneously identifying, measuring, and analyzing activity from individual cells within a neural population in live brain tissue. Analyzing fluorescence microscopy (FM) images in real-time with computational algorithms is essential for achieving this goal. Deep learning techniques have shown promise in this area, but face domain-specific challenges due to limited training data, significant amounts of voxel noise in FM images, and thin structures present in large 3D images. In this thesis, I address these issues by introducing a novel deep learning pipeline to analyze static FM images of neurons with minimal data requirements and demonstrate the pipeline’s ability to segment neurons from low signal-to-noise ratio FM images with few training samples. The first step of this pipeline employs a Generative Adversarial Network (GAN) equipped to learn imaging properties from a small set of static FM images acquired for a given neuroscientific experiment. Operating like an actual microscope, our fully-trained GAN can then generate realistic static FM images from volumetric reconstructions of neurons with added control over the intensity and noise of the generated images. For the second step in our pipeline, a novel segmentation network is trained on GAN-generated images with reconstructed neurons serving as “gold standard” ground truths. While training on a large dataset of FM images is optimal, a 15\% improvement in neuron segmentation accuracy from noisy FM images is shown when architectures are fine-tuned only on a small subsample of real image data. To evaluate the overall feasibility of our pipeline and the utility of generated images, 2 novel synthetic and 3 newly acquired FM image datasets are introduced along with a new evaluation protocol for FM image ”realness” that incorporates content, texture, and expert opinion metrics. While this pipeline's primary application is to segment neurons from highly noisy FM images, its utility can be extended to automate other FM tasks such as synapse identification, neuron classification, or super-resolution.</p> |
| first_indexed | 2024-09-25T04:10:15Z |
| format | Thesis |
| id | oxford-uuid:40d4b805-a9df-4371-8ce1-1080653d6cab |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:10:15Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:40d4b805-a9df-4371-8ce1-1080653d6cab2024-06-19T14:00:07ZFluorescence microscopy image analysis of retinal neurons using deep learningThesishttp://purl.org/coar/resource_type/c_db06uuid:40d4b805-a9df-4371-8ce1-1080653d6cabEnglishHyrax Deposit2023Cudic, MNoble, AGrau, VDiamond, JDmitri, R<p>An essential goal of neuroscience is to understand the brain by simultaneously identifying, measuring, and analyzing activity from individual cells within a neural population in live brain tissue. Analyzing fluorescence microscopy (FM) images in real-time with computational algorithms is essential for achieving this goal. Deep learning techniques have shown promise in this area, but face domain-specific challenges due to limited training data, significant amounts of voxel noise in FM images, and thin structures present in large 3D images. In this thesis, I address these issues by introducing a novel deep learning pipeline to analyze static FM images of neurons with minimal data requirements and demonstrate the pipeline’s ability to segment neurons from low signal-to-noise ratio FM images with few training samples. The first step of this pipeline employs a Generative Adversarial Network (GAN) equipped to learn imaging properties from a small set of static FM images acquired for a given neuroscientific experiment. Operating like an actual microscope, our fully-trained GAN can then generate realistic static FM images from volumetric reconstructions of neurons with added control over the intensity and noise of the generated images. For the second step in our pipeline, a novel segmentation network is trained on GAN-generated images with reconstructed neurons serving as “gold standard” ground truths. While training on a large dataset of FM images is optimal, a 15\% improvement in neuron segmentation accuracy from noisy FM images is shown when architectures are fine-tuned only on a small subsample of real image data. To evaluate the overall feasibility of our pipeline and the utility of generated images, 2 novel synthetic and 3 newly acquired FM image datasets are introduced along with a new evaluation protocol for FM image ”realness” that incorporates content, texture, and expert opinion metrics. While this pipeline's primary application is to segment neurons from highly noisy FM images, its utility can be extended to automate other FM tasks such as synapse identification, neuron classification, or super-resolution.</p> |
| spellingShingle | Cudic, M Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title | Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title_full | Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title_fullStr | Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title_full_unstemmed | Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title_short | Fluorescence microscopy image analysis of retinal neurons using deep learning |
| title_sort | fluorescence microscopy image analysis of retinal neurons using deep learning |
| work_keys_str_mv | AT cudicm fluorescencemicroscopyimageanalysisofretinalneuronsusingdeeplearning |