Deep Representation Learning with Limited Data for Biomedical Image Synthesis, Segmentation, and Detection
Loading...
Authors
Kamran, Sharif Amit
Issue Date
2023
Type
Dissertation
Language
Keywords
Biomedical Imaging , Deep Learning , Generative Adversarial Networks , Image Segmentation , Object Tracking , Representation Learning
Alternative Title
Abstract
Biomedical imaging requires accurate expert annotation and interpretation that can aid medical staff and clinicians in automating differential diagnosis and solving underlying health conditions. With the advent of Deep learning, it has become a standard for reaching expert-level performance in non-invasive biomedical imaging tasks by training with large image datasets. However, with the need for large publicly available datasets, training a deep learning model to learn intrinsic representations becomes harder. Representation learning with limited data has introduced new learning techniques, such as Generative Adversarial Networks, Semi-supervised Learning, and Self-supervised Learning, that can be applied to various biomedical applications. For example, ophthalmologists use color funduscopy (CF) and fluorescein angiography (FA) to diagnose retinal degenerative diseases. However, fluorescein angiography requires injecting a dye, which can create adverse reactions in the patients. So, to alleviate this, a non-invasive technique needs to be developed that can translate fluorescein angiography from fundus images. Similarly, color funduscopy and optical coherence tomography (OCT) are also utilized to semantically segment the vasculature and fluid build-up in spatial and volumetric retinal imaging, which can help with the future prognosis of diseases. Although many automated techniques have been proposed for medical image segmentation, the main drawback is the model's precision in pixel-wise predictions. Another critical challenge in the biomedical imaging field is accurately segmenting and quantifying dynamic behaviors of calcium signals in cells. Calcium imaging is a widely utilized approach to studying subcellular calcium activity and cell function; however, large datasets have yielded a profound need for fast, accurate, and standardized analyses of calcium signals. For example, image sequences from calcium signals in colonic pacemaker cells ICC (Interstitial cells of Cajal) suffer from motion artifacts and high periodic and sensor noise, making it difficult to accurately segment and quantify calcium signal events. Moreover, it is time-consuming and tedious to annotate such a large volume of calcium image stacks or videos and extract their associated spatiotemporal maps. To address these problems, we propose various deep representation learning architectures that utilize limited labels and annotations to address the critical challenges in these biomedical applications. To this end, we detail our proposed semi-supervised, generative adversarial networks and transformer-based architectures for individual learning tasks such as retinal image-to-image translation, vessel and fluid segmentation from fundus and OCT images, breast micro-mass segmentation, and sub-cellular calcium events tracking from videos and spatiotemporal map quantification. We also illustrate two multi-modal multi-task learning frameworks with applications that can be extended to other domains of biomedical applications. The main idea is to incorporate each of these as individual modules to our proposed multi-modal frameworks to solve the existing challenges with 1) Fluorescein angiography synthesis, 2) Retinal vessel and fluid segmentation, 3) Breast micro-mass segmentation, and 4) Dynamic quantification of calcium imaging datasets.
Description
Citation
Publisher
License
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 United States