Syer, Tom;
(2025)
Evaluation of Machine Learning Approaches for Classification of Prostate Cancer on Magnetic Resonance Imaging.
Doctoral thesis (Ph.D), UCL (University College London).
Abstract
MRI is now well-established for the detection of suspicious prostate lesions before biopsy for the diagnosis of localised prostate cancer, and a normal MRI can often result in the safe omission of biopsy. There is still potential for improvements to the diagnostic pathway and its accuracy, particularly to reduce the number of unnecessary biopsies. Machine learning has been hailed as having the potential to improve the diagnostic accuracy, potentially surpassing that of clinical radiologists. However, relatively few examples of machine learning are used in clinical practice, particularly for aiding diagnosis. Therefore, this thesis aims to assess the current issues preventing machine learning translation into clinical practice and investigate potential methodologies and infrastructure to enable robust evaluation and inform future studies. I initially conducted a systematic review of the literature investigating machine learning applications using MRI for detecting and classifying prostate cancer, assessing study design quality and the common limitations that may prevent clinical translation. Following this, in Chapter 3, I externally evaluated a promising logistic regression model in a multi-reader setting in comparison to commonly used clinical thresholds. I assessed the impact of contour variation and choice of slice and timepoint for dynamic contrast-enhanced (DCE) images between readers on the model predictors and outcome. I then investigated the impact of different updating methods to improve the model’s performance, calibration and generalisability in Chapter 4 as well as the likely stability of these updated models in Chapter 5. In Chapter 6, I describe the setup and implementation of an AI challenge allowing for independent evaluation of machine-learning classifiers for prostate lesion classification with a large, multicentre, deeply phenotyped cohort and present the results of one submission. In Chapter 7, I describe the preparation of a multi-centre prostate MRI dataset made available publicly and subsequent external evaluation of a whole patient classifier model that did not require radiologist contours with this dataset. I investigated its potential use in reducing radiologist reporting requirements and biopsy rates and the impact of updating the threshold and logistic calibration with varying sizes calibration sets to help maintain the desired sensitivity and improve calibration. Finally, I discuss the overall conclusions from this thesis and its impact on future work.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Evaluation of Machine Learning Approaches for Classification of Prostate Cancer on Magnetic Resonance Imaging |
| Language: | English |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Medical Sciences UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Medical Sciences > Div of Medicine |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10212776 |
Archive Staff Only
 |
View Item |
