Saeed, Shaheer Ullah; (2024) Multi-level optimisation using deep meta learning for medical image analysis. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhD_Dissertation_revised.pdf - Other Download (11MB) | Preview

Abstract

Performance of medical imaging tasks is strongly associated with image quality. Imaging defects, may, however, impact different tasks differently e.g., an artefact covering a lesion in the prostate may hinder cancer detection, however, even if a portion of the gland is unaffected by the artefact, the image may still be usable for gland presence classification. Unlike most general quality assessment techniques that quantify quality based on proportion of artefacts, in a task-indifferent manner, task-specific quality directly measures the sample’s impact on a downstream task. We propose to use the task performance of machine learning algorithms as an objective measure of task-specific quality. This is learnt using bi-level optimisation involving two functions: 1) a task predictor performing the downstream clinical task, learnt using a training set; and 2) a controller predicting task-specific quality, learnt as a weighting policy by maximising a reward based on the task predictor’s performance, measured on a controller-weighted validation set. Training the two simultaneously leads to the controller learning to weight samples with a negative impact on task performance lower, and the task predictor learning to perform the downstream task. This mechanism leads to a definition of task-specific quality informed directly by the task performance as opposed to any subjective human labels of quality. Firstly, we develop a meta-learning methodology to learn this task-specific quality (Chapter 2) and propose to learn it using reinforcement learning (Chapter 3). Secondly, We equip the system with the ability to detect the underpinning causes of low task-specific quality, by distinguishing imaging-defect-impacted vs clinically challenging (e.g., unusual pathology presentation) samples (Chapter 4). Thirdly, we propose mechanisms to enable improved adaptability to new tasks with few labelled samples for the task, using multi-level meta-learning (Chapter 5). We demonstrate the efficacy for a variety of clinical tasks and imaging modalities in the relevant chapters. Finally, we extend the methodology for a relevant active learning (training-data prioritisation) scenario, for performance convergence with fewer labelled samples compared to the common random prioritisation (Chapter 6). This active learning can be considered as a similar problem where the impact of adding training data samples needs to be measured in order to indicate potential performance gain by the usage of the sample for training.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item