Abstract
This thesis presents a dual exploration of machine learning applications in predictive healthcare, focusing on prostate Magnetic Resonance (MR) image segmentation and intradialytic hypotension (IDH) prediction in haemodialysis patients. The research is motivated by the need for precise diagnostic tools and real-time intervention capabilities in urology, with the overarching aim of improving patient outcomes through advanced, data-driven methods.Traditional diagnostic pathways for prostate cancer primarily rely on prostate- specific antigen (PSA) testing and transrectal ultrasound (TRUS)-guided biopsies. PSA screening has a low specificity rate of approximately 36%, leading to false positives. MR imaging emerges as a promising alternative. In prostate MR image segmentation, convolutional neural networks (CNNs) were employed to develop seg- mentation models, with the two-stage end-to-end MultiResUNet approach achieving a Dice Similarity Coefficient (DSC) of 0.804 and a Hausdorff Distance (HD) of 4.818. These results demonstrate the model’s capacity for high anatomical accuracy, es- sential for prostate cancer diagnosis and treatment planning. Comparative analysis revealed that multi-stage models significantly outperformed single-stage segmenta- tion approaches, underscoring the advantages of progressive refinement in handling complex anatomical structures.
End-stage renal disease patients can undergo haemodialysis at home or in clini- cal settings to sustain life. However, a common and challenging complication of haemodialysis is IDH, which can lead to various medical consequences, including patient discomfort, dialysis interruption, hospitalisation, increased morbidity, mor- tality, cerebral ischaemia, vascular access thrombosis, and cardiovascular events. For IDH prediction, this study investigated multiple machine learning models, in- cluding Random Forest (RF), Long Short-Term Memory (LSTM), and a Selective Stacked Ensemble Learning (SSEL) model, with SSEL achieving the highest Re- ceiver Operating Characteristic - Area Under the Curve (ROC-AUC) of 0.7209. Key physiological predictors, such as pre-dialysis blood pressures, were identified, validating the model’s relevance for clinical applications. The findings suggest that incorporating real-time and longitudinal patient data could further enhance predic- tive accuracy, offering potential for proactive interventions in dialysis care to reduce adverse outcomes.
This thesis addresses several challenges, such as class imbalance, dataset variability, and computational demands, proposing future research directions that include adap- tive architectures, multi-source training datasets, and real-time data integration. Collectively, these findings establish a framework for leveraging machine learning in clinical settings, positioning artificial intelligence as a critical tool in person- alised healthcare for improving diagnostic precision and enhancing patient safety in prostate cancer and haemodialysis care.
Keywords: Image segmentation, prostate, magnetic resonance imaging, intradia- lytic hypotension
| Date of Award | 23 May 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Mohamed Bader-El-Den (Supervisor), Vincenzo Tamma (Supervisor) & Shamsul Masum (Supervisor) |