Data Decomposition Methods for Medical Image Classification

Husien, Asmaa (2025) Data Decomposition Methods for Medical Image Classification. Doctoral thesis, Birmingham City University.

Preview

Text Asmaa Husien PhD _Thesis_Final Version_Final Award June 2025.pdf - Accepted Version Download (28MB)

Abstract

Although deep learning methods have achieved outstanding success in the medical image field, they face several challenges that can significantly impact training effectiveness, learning stability and meaningful generalisations. One of these challenges is the limited availability of annotated samples for certain diseases, which are often difficult and expensive to obtain. Another important issue is the presence of overlapping class distributions, where similarities between features of different classes make it difficult for the model to distinguish between them accurately.

To address these issues, the thesis aims to develop deep learning solutions that improve classification performance when faced with limited annotations and irregular class distributions. The study specifically focuses on three key objectives: (1) designing a convolutional neural network that improves feature learning from a generic domain to a more specific task with small annotated samples, (2) developing a deep learning model that effectively mitigates class overlap by refining class boundaries, and (3) enhancing the generalisation strategy to improve learning stability and simplify complex patterns within datasets.

To achieve these objectives, the thesis presents three main contributions. First, the 4S-DT model and its advanced version, XDecompo, are introduced to enhance feature transferability through self-supervised learning with sample decomposition and overcome the limited samples of the dataset. 4S-DT uses the k-means clustering to perform the decomposition process, which may not always align with the true structure of the data. In contrast, XDecompo employs an affinity propagation-based class decomposition to automatically enhance the learning of the class boundaries in the downstream task without the need for preset cluster numbers. This clustering process provides more flexibility and adaptability compared to the parametric approach used by 4S-DT. Moreover, XDecompo also incorporates an explainable component to highlight salient pixels that influenced the model’s decision and explain the effectiveness of XDecompo to enhance the feature extraction and increase the trust of deep learning applications.

The second contribution introduces CLOG-CD, a convolutional neural network that integrates curriculum learning with class decomposition to improve classification performance on medical image datasets exhibiting class irregularities. CLOG-CD also explores different oscillation steps to evaluate the impact of varying learning speeds on model generalisation at different levels of granularity.

The third contribution of the thesis introduces a novel curriculum learning with a progressive of self-supervised learning called (CURVETE) that employs a curriculum learning strategy based on the granularity of sample decomposition during the training of unlabelled samples. Through this process, CURVETE enhances the quality of feature representations, extracting rich information across different levels of granularity. These features can then be effectively transferred to a new downstream task with limited examples, ultimately improving classification performance. CURVETE also handles the challenge of irregular class distribution by utilising the curriculum learning strategy with a class decomposition approach in the downstream task.

Extensive experiments have been carried out on various medical image datasets, utilising different evaluation metrics, to validate the effectiveness of our three contributions to the thesis. For the first contribution, 4S-DT has achieved a high accuracy of 97.54% and 99.80% for detecting COVID-19 cases in dataset-A and dataset-B, respectively. Additionally, XDecompo achieved accuracies of 96.16% and 94.30% for colorectal cancer and brain tumour images, respectively, outperforming 4S-DT and other training strategies. The second contribution, CLOG-CD, achieved an accuracy of 96.08% on the chest x-ray dataset, 96.91% on the brain tumour dataset, 79.76% on the digital knee x-ray, and 99.17% on the colorectal cancer dataset using the baseline ResNet-50. In addition, CLOG-CD using DenseNet-121 achieved 94.86%, 94.63%, 76.19%, and 99.45% for chest x-ray, brain tumour, digital knee x-ray, and colorectal cancer datasets, respectively. Finally, CURVETE showed significant improvements in performance with an accuracy of 96.60% on the brain tumour dataset, 75.60% on the digital knee x-ray dataset, and 93.35% on the Mini-DDSM dataset using the baseline ResNet-50. Furthermore, the classification performance with the baseline DenseNet-121 achieved an accuracy of 95.77%, 80.36%, and 93.22% on the brain tumour, digital knee x-ray, and Mini-DDSM datasets, respectively.

Item Type:	Thesis (Doctoral)
Dates:	Date Event 9 June 2025 Accepted
Uncontrolled Keywords:	Data Decomposition, medical image classification
Subjects:	CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science CAH11 - computing > CAH11-01 - computing > CAH11-01-05 - artificial intelligence CAH11 - computing > CAH11-01 - computing > CAH11-01-08 - others in computing
Divisions:	Doctoral Research College > Doctoral Theses Collection Faculty of Computing, Engineering and the Built Environment > College of Computing
Depositing User:	Louise Muldowney
Date Deposited:	01 Jul 2025 09:01
Last Modified:	01 Jul 2025 09:01
URI:	https://www.open-access.bcu.ac.uk/id/eprint/16454

Actions (login required)

View Item