Automatic Diagnosis of Coronavirus Disease on Chest CT Images Utilising Deep Learning

The ongoing pandemic, Coronavirus disease 2019 (COVID-19), caused by the SARS-CoV-2 virus, has presented substantial challenges, particularly in the realm of diagnostic accuracy and the labour-intensive nature of the work of radiologists. While computer-aided systems have shown promise as potential solutions, several challenges persist. These include issues related to performance sensitivity, limited training data, high annotation costs, and resource efficiency demands. This thesis aims to address these challenges and inform the development of highly sensitive, data-efficient deep neural networks (DNNs). The research in this thesis evolved through three distinct stages, each driven by changing research needs. Initially, due to limited public availability of large-scale data sources, focus was placed on classification with a restricted dataset. GBBNet was proposed by reconfiguring and retraining the bestselected backbone, GoogLeNet, based on a prior optimal-backbone-selection algorithm. This model was specifically designed for COVID-19 diagnosis, leveraging pre-trained representations. And Bayesian Optimization was employed for robust hyperparameter tuning. A significant milestone during this phase was the development of the COVIDSeg model, designed for precise lung area delineation. In the second stage, the research transitioned from singular pre-trained to ensemble learning-based models. AdaD-FNN was presented to leverage sequentially transferred knowledge and decay mechanisms to reduce noise retention and enhance pattern recognition in complex cases. Acknowledging the heavy human involvement in the COVID-Seg model, especially in atypical cases, the F-U2MNet-C segmentation model was designed. This advanced preprocessing tool demonstrated superior generalisability, countering issues such as block effects, disparity in enhancement levels, and boundary ambiguity. Recognising the resource-intensive nature of ensemble learning, a fuzzy logic Proportional-Integral- Derivative (PID) Controller-Based Stochastics Optimization algorithm with an adaptive learning rate was proposed. This method alleviated the early overshoot phenomenon observed with Stochastic Gradient Descent-Momentum (SGD-M) and expedited DNN convergence. The efficacy of this algorithm is evident in its ability it harnesses past, present, and anticipated changes of the gradient, updating network parameters in line with specific fuzzy logic rules. Experimental results underline the importance of this optimizer in addressing early overshooting and promoting better convergence, holding promise for enhanced clinical diagnosis. After conducting thorough testing on publicly available Chest Computed Tomography (CCT) image datasets, these three frameworks demonstrated superior performance compared to stateof- the-art methods in diagnosing COVID-19, resulting in significant enhancements in both accuracy and robustness.