Deep Learning-Based Decision Support for Multi-Phase Cancer Management
Basaad, Abdullah Ahmed (2025) Deep Learning-Based Decision Support for Multi-Phase Cancer Management. Doctoral thesis, Birmingham City University.
Preview |
Text
Abdullah Ahmed Basaad PhD Thesis_Final Version_Final Award November 2025.pdf - Accepted Version Download (16MB) |
Abstract
Early and accurate prediction of cancer is important for improving patient outcomes and making appropriate treatment decisions. This research will explore how deep learning technologies can assist clinicians in treating multi-stage cancers, specifically lung and breast cancer. The first phase is the early stage detection of lung cancer through diagnostic imaging and generating reports. Diligent follow-up care of patients, while very important, is also a big part of ongoing patient care through routine check ups and monitoring tools such as mammograms, MRIs, and tumour markers, which fall outside of the scope of this research. The second phase concerns breast cancer recurrence; breast cancer often returns as local or regional recurrence during follow-up with the intent that it can be treated curatively. The third phase considers metastasis, which often manifests later and is a more serious event. In some cases, particularly with aggressive cancers or advanced stage cancers, metastasis may be the first sign of recurrence, but typically it comes following local recurrence.
First we consider the possibility of using convolutional neural networks (CNNs) to identify lung cancer in chest X-ray images early in the disease process to avoid treatment delays and improve the odds of survival. The CNNs are used for generating medical reports, but we also want to incorporate data from clinical patient assessments and observations. A secondary focus of this study is to develop predictive models which estimate the statistical likelihood of relapse in breast cancer to assist with treatment planning and diagnostics. Third, we want to outline a deep learning strategy for predicting the risk of metastasis in breast cancer patients, to alert clinicians to the risk of progression of the disease, which could have implications for the clinician’s treatment decision. Overall our analysis suggested that deep learning has merit as a game-changer for early detection and the long-term management of our patients with complex cancers. In this thesis, three novel methods are presented to improve the decision supporting for multi-phase cancer management.
Initially, AIM-X Attention-Infused Multimodal Cross-Interaction for X-ray Clinical Report Generation is an innovative AI-powered diagnosis system which combines advances in CAD systems, including multimodal imaging analysis and radiology report generation, to enhance clinical decision-making. The system combines an attention mechanism and multiscale feature extraction with hyperparameter optimisation using a genetic algorithm, thus refining the diagnostic accuracy. It is designed to incorporate radiological images in addition to medical reports in a textual format achieving an accuracy up to 94.5%.
Secondly, GraphX-Net, a GNN based on Shapley Value, is proposed for predicting cancer recurrence by modelling patients as nodes in a graph and using a set of clinical factors such as tumour cellularity and hormone therapy. With the usage of graph convolutional layers in conjunction with Shapley values, this model efficiently evaluates patient attributes and patient-to-patient relations, demonstrating best performances in recurrence prediction with 98% accuracy and 98.4% F1-score.
Third, the BG-MBC BERT-GNN approach for the prediction of metastatic breast cancer using histopathological reports, which combines LLM and GNN, is formulated for the diagnosis of MBC. This is a blend of natural language processing and graph learning paradigms, in which semantic understanding is extracted using BERT embeddings of pathology reports and understanding of patient relations is enabled using GNN guided by attention scores from the embeddings. The model achieves an accuracy of 98% and 99% accuracy in cross-validation, demonstrating strong predictive performance.
These collectively contribute to a developing field of AI-based medical insight that allows robust solutions toward disease detection in early stages of lung cancer and enhanced breast cancer recurrence and metastatic prediction.
Actions (login required)
 |
View Item |
Tools
Tools
