Biomedical Engineering - Theses
Permanent URI for this collection
Search Results
Now showing
1 - 1 of 1
-
ItemNo Preview AvailableNovel machine-learning approaches to create a structurally accurate virtual model of the heart cellKhadankishandi, Afshin ( 2023-06)Cardiomyocytes are densely packed with parallel columns of myofibrils and mitochondria. Research has shown a strong correlation between changes in ultrastructure and changes in the heart’s function. For example, heart pumping is often compromised in heart diseases such as cardiomyopathy, hypertension and diabetes. Hence, understanding the 3D architecture of cardiac cells will underpin breakthroughs in cardiovascular disease treatment and prevention. With the advent of high-throughput microscopy image datasets resulting from modalities such as serial block-face and focused ion-beam scanning electron microscopy, we can acquire large datasets of cardiac muscle cells in 3D. However, segmenting these datasets is challenging due to low contrast and high noise ratio. The community often relies on manual segmentation and image tracing, a laborious and cost-inefficient approach that hinders novel breakthroughs. This thesis proposes state-of-the-art deep neural networks to segment ultrastructures of cardiac cells in EM datasets and obtains 3D statistical architecture of the cardiomyocyte. EM-net and EM-stellar are cloud-based software proposed to segment EM image datasets and benchmark a wide range of segmentation performance measures. EM-net is a scalable convolutional neural network offering fast convergence during optimisation and can be trained with minimal ground-truth information due to its novel architecture. EM-stellar is hosted on Google Colab, and it can be used to benchmark the performance of state-of-the-art deep neural networks on a user-specified dataset. Together these pipelines offer the research community more efficient ways to segment and analyse cardiac muscle ultrastructure from electron microscopy datasets. Finally, we propose CardioVinci, a workflow utilising generative adversarial networks (GANs) to obtain a statistical 3D model of cardiomyocyte architecture. CardioVinci addresses a significant challenge with large EM datasets: the time taken to collect tissue samples, acquire the data, extract key characteristics and statistically analyse 3D changes in the ultrastructures. It encodes the 2D and 3D variations in the ultrastructures across the image volume into a generative model. As a result, the community will be able to statistically quantify the morphologies and spatial assembly of mitochondria, myofibrils, and Z-disks with minimal manual annotation.