Tools
Tools
Bohoran, TA, 2024. Streamlining clinical pipelines for cardiovascular imaging with deep learning. PhD, Nottingham Trent University.
Preview |
Text
Tuan_Bohoran_2025.pdf - Published version Download (6MB) | Preview |
Abstract
Cardiovascular diseases continue to be the leading cause of morbidity and mortality worldwide, emphasising the need for accurate diagnosis, efficient monitoring, and timely intervention in managing these conditions. This thesis seeks to address key challenges in the cardiovascular field by proposing novel, resource-efficient computational techniques aimed at improving the accuracy and reliability of cardiovascular assessments through the application of cutting-edge machine learning (ML) and deep learning (DL) approaches.
The first theme introduces a new DL model that automatically segments the aortic lumen from cine cardiovascular magnetic resonance (CMR) images. The model, based on bi-directional ConvLSTM (BConvLSTM) U-Net with densely connected convolutions, provides a fresh perspective on measuring aortic distensibility (AD). It addresses significant challenges in existing methods by using a hierarchical learning framework that efficiently processes spatio-temporal aspects of video inputs. The model combines encoder and decoder feature maps through non-linear functions, and manages the high class imbalance in the data through using an appropriate loss function. The study succeeded in applying the model to a multi-centre, multi-vendor dataset with diverse patient demographics. The results show that the proposed model exceeds the current state-of-the-art methods in accuracy. Moreover, it achieves this with significantly less environmental impact, consuming approximately ∼3.9 times less fuel and generating ∼2.8 times fewer carbon emissions. This model has great potential as a tool for exploring genome-wide associations between AD, aortic areas, and cognitive performance in expansive biomedical databases. The study’s focus on energy usage and carbon emissions demonstrates the commitment to sustainable deep learning practices. This research not only enhances the capabilities of CMR-derived aortic stiffness evaluation but also sets the stage for more widely applicable and systematic deep learning-powered methodologies in medical imaging, balancing accuracy with environmental consciousness.
In the ongoing endeavour to predict right ventricular (RV) volume from 2D echocardiography images, there is a lack of accurate methods that take advantage of planimetry data. The second theme delves into this intricate task. It analyses 100 RV volumes, and tabular input information encompasses planimetry data from eight standard echocardiography views, as well as age, gender, and cardiac phase information. We present two ML regression methods herein. The first method utilises a gradient-boosted regression tree (GBRT)-based method, enhanced by a supervised tree kernel, in order to not only forecast RV volume but also estimate uncertainty in predictions. The second method employs a multi-head attention-based transformer (called the feature tokeniser transformer), which tokenises tabular data by distinguishing between numerical and categorical inputs. Our findings indicate that while the initial GBRT-based method shows promise, it exhibits limitations in prediction accuracy. There is a significant improvement in RV volume prediction accuracy with the transformer-based model, surpassing the initial GBRT-based approach. This research highlights the importance of incorporating attention mechanisms and feature tokenising in methodological strategies. Additionally, we conduct a "gain" explainability analysis using the GBRTs, facilitating the development of more clinically viable pipelines. This will ultimately lead to more refined and accurate predictive models in RV volume analysis based on echocardiography.
Arrhythmia, characterised by irregular heartbeats, poses significant health risks. Residual networks, a subset of DL architectures, have emerged as a potent tool to detect abnormalities in electrocardiogram signal anomalies. However, the enhanced accuracy and capabilities afforded by increasing network depth in these models come at the cost of heightened computational demands, which poses a considerable challenge to their practical applicability. Addressing this critical bottleneck, the third theme presents a methodology for the resource-economical development of ML-enabled systems for arrhythmia detection. The proposed methodology, grounded in the dynamical system perspective of residual networks, initiates the training process with a shallow network and then progressively increases its depth. We rigorously validate the method on the PhysioNet MIT-BIH arrhythmia data set using heartbeat spectrograms as training input. The results show that the proposed training requires a minimum of 39.47% fewer parameters per epoch compared to conventional vanilla training, a feat achieved without sacrificing and potentially improving overall performance. Our findings suggest the methodology not only drastically reduces training time but also promises significant savings in energy consumption and environmental costs, offering a glimpse into a future of more sustainable and resource-efficient machine learning developments in arrhythmia detection.
Semantic segmentation enables a higher level of understanding of visual scenes. It also forms an essential capability towards delivering a plethora of life-changing technologies. However, the discrepancy between the distribution of the input samples used to train the model and the input distribution encountered during testing (commonly known as covariate shift), as well as the lack of trustworthy methods for detecting it, hinder the practical application of semantic segmentation. In the fourth theme, we develop a reliable, sample-efficient, distribution-free and model-agnostic hypothesis test, named the Segmetron, to detect detrimental covariate shift in semantic segmentation. To deal with the intractability of the above problem and deliver strong performance guarantees on unknown arbitrary test distributions, we build upon recent advancements in the PQ learning setting of selective classification, and extend it to a different discriminative model (i.e. segmenters). To assess an unlabelled target domain, Segmetron leverages an existing (but random) pre-trained semantic segmentation model and the labelled samples used to train it. To train the enforced disagreement segmenters to learn the same generalisation region as the pre-trained semantic segmentation model, we propose loss functions (to agree) which are more apropos to the semantic segmentation task. Our approach stands apart from previous studies on semantic segmentation robustness which relied on synthetic domain shifts. Instead, we analyse two real-world covariate shifts from the cardiovascular magnetic resonance imaging field, concerned with binary (aorta, background) and multi-class (left ventricle, right ventricle, myocardium, background) semantic segmentation tasks. Our experiments demonstrate that the Segmetron hypothesis test outperforms other state-of-the-art techniques in terms of statistical power on both semantic segmentation tasks, given access to only a 3D dataset from one patient. This work holds considerable value because it aligns with “Responsible AI” principles and it happens at a time when the machine learning community is striving to increase public trust and acceptance of AI technologies. Moreover, Segmetron has the potential to support the successful deployment of a plethora of semantic segmentation-based transformative AI solutions.
| Item Type: | Thesis |
|---|---|
| Creators: | Bohoran, T.A. |
| Contributors: | Name Role NTU ID ORCID |
| Date: | October 2024 |
| Divisions: | Schools > School of Science and Technology |
| Record created by: | Jeremy Silvester |
| Date Added: | 12 Dec 2025 12:08 |
| Last Modified: | 12 Dec 2025 12:08 |
| URI: | https://irep.ntu.ac.uk/id/eprint/54865 |
https://orcid.org/0000-0001-7403-923XActions (login required)
 |
Edit View |
Statistics
Views
Views per month over past year
Downloads
Downloads per month over past year
