Thesis
Application of Meta Pseudo Labels for semantic segmentation of axons within optic nerve cross sections
University of Iowa
Master of Science (MS), University of Iowa
Spring 2023
DOI: 10.25820/etd.007140
Abstract
Deep learning methods for medical imaging often face challenges due to limited and difficult-to-produce ground truth images for training. This thesis investigates the effects of a semi-supervised training model called Meta Pseudo Labels (MPL) for the semantic segmentation of axons within optic nerve cross-sectional images. The MPL training model is integrated with the AxonDeep feature pyramid network architecture, previously employed for axon segmentation in optic nerve cross-section images. To evaluate MPL's performance with limited labeled training data, four model versions are trained using 10%, 25%, 50%, and 100% of the available labeled data, replicating the experiments conducted with AxonDeep.
The same training, validation, and testing data used in the original AxonDeep experiments are employed to ensure a fair comparison. MPL operates in a semi-supervised manner, leveraging both labeled and unlabeled data, and employs self-supervised learning principles by generating pseudo labels for unlabeled images to enhance the training process. The results show that MPL can effectively utilize minimal training data to achieve an accurate segmentation, with the 10% MPL model yielding a similar Dice score to the 100% generative adversarial network (GAN) AxonDeep model, though the axon counts do not achieve better results than AxonDeep (GAN).
Details
- Title: Subtitle
- Application of Meta Pseudo Labels for semantic segmentation of axons within optic nerve cross sections
- Creators
- Ashelyn Mann
- Contributors
- Mona K. Garvin (Advisor)Michael G. Anderson (Committee Member)Tyler Bell (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Spring 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007140
- Number of pages
- x, 62 pages
- Copyright
- Copyright 2023 Ashelyn Mann
- Language
- English
- Date submitted
- 04/25/2023
- Date approved
- 05/10/2023
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 56-62).
- Public Abstract (ETD)
Deep learning, a type of artificial intelligence (AI), can help analyze medical images, such as those of the optic nerves. However, these methods often require large amounts of labeled images that are already thoroughly analyzed to train the AI. This data can be difficult to obtain, so methods that require less pre-analyzed data to train AI are pursued. This thesis explores a semi-supervised training model called Meta Pseudo Labels (MPL) that can use a few labeled images and a larger amount of unlabeled images to learn and improve its capabilities.
MPL is combined with an existing method called AxonDeep to segment axons, the long part of nerve cells that transmit information, in optic nerve cross-section images. To see how well MPL works with access to less labeled data, we train four versions of the model with varying amounts of labeled data (10%, 25%, 50%, 100%).
The results show that MPL can effectively use minimal pre-analyzed training data to accurately segment axons in optic nerve images. For example, the model trained with only 10% of the labeled data performs similarly at segmenting axons to AxonDeep when it used 100% of the labeled data. However, MPL was not as accurate as AxonDeep at providing the correct number of axons in optic nerve images.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984425389402771
Metrics
19 Record Views