Abstract
Deep learning techniques have recently achieved impressive success in medical diagnosis and physiological signal processing. This research aims to improve the classification performance of convolutional neural networks (CNNs) by means of employing a dynamic weight structure, modifying the architectures of existing CNN models, and combining different machine learning algorithms which aims to alleviate the difficulty in model training. The proposed methods have been applied to the medical signal processing and image classification, such as epileptic seizures phases classification, hand gesture recognition, and COVID-19 detection.The main contributions of this research include: 1) proposing novel deep learning architectures which can improve the classification performance, generalisation ability, and robustness of the existing deep learning techniques; 2) establishing a classification system which consists of automatic representation learning, dimensionality reduction, and classification process; 3) combining supervised learning and unsupervised learning algorithms; 4) designing a two-step classification system and improving the fuzzy c-means clustering method; 5) developing a type of dynamic CNN modification method which is able to alleviate the overfitting and vanishing gradient problems; 6) applying the proposed algorithms to the classification of medical signals and images. Especially, the proposed methods have achieved state-of-the-art performance in the classification of epileptic seizure types with electroencephalographic (EEG) signals, the classification of hand gestures with electromyographic (EMG) signals, and the classification of COVID-19 cases with chest X-Ray (CXR) and computed tomography (CT) images. Specifically, the proposed methodology and outcomes of this research are as follows.
Firstly, a type of variable weight convolutional neural networks (VWCNN) is proposed, in which the weights of kernels can change adaptively according to the input data. Differing from conventional CNNs whose weights are static after training, the proposed VWCNNs incorporates the training of conventional CNNs and a parameter tuning structure, which can be viewed as an infinite number of traditional, fixed-weight CNNs. The forward and backward propagations of the proposed method are given in mathematical formulas. The proposed VWCNNs are applied to an epileptic seizure phase dataset and a benchmark EEG dataset of seven seizure types. The proposed method achieves 100% test accuracy and shows strong robustness in epileptic seizure phase classification.
Secondly, two classification frameworks are proposed which aims to combine the supervised and unsupervised learning algorithms. The first framework is a combination of convolutional autoencoder stacks and CNN, which is applied to the classification of hand gestures with EMG signals. The second framework employs clustering algorithms and CNN in the classification of hand gestures. In the second framework, the fuzzy c-means clustering algorithm is generalised to be able to deal with multi-class clustering tasks. The proposed methods achieve 100% classification accuracy and strong robustness in the classification of ten hand gestures. Also, a federated learning framework is proposed based on transfer learning methods, which is used to train personalised local models for distributed users.
Thirdly, a type of CNN modification method is designed to solve the vanishing gradient problem and improve the classification performance. The modified MobileNet is applied to the classification of COVID-19, Tuberculosis, viral pneumonia (with the exception of COVID-19), bacterial pneumonia and normal controls with chext X-Ray images. Moreover, the proposed modified ResNet is used for the classification of COVID-19, non-COVID-19 infections and normal controls with CT images. The results show that the proposed methods achieve 99.6% test accuracy on the five-category CXR image dataset and 99.3% test accuracy on the CT image dataset. Six advanced CNN architectures and two baseline models for COVID-19 detection are employed for comparison. Two benchmark datasets and a CXR image dataset which combines eight different CXR image sources are employed to evaluate the performance of the above models. The results show that the proposed methods outperform the comparative models in classification accuracy, sensitivity, and precision, which demonstrate their potential in computer-aided diagnosis for healthcare applications.
| Date of Award | 1 Dec 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Hak-Keung Lam (Supervisor) & Hongbin Liu (Supervisor) |