Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.815277
Title: Reliable machine learning for medical imaging data through automated quality control and data harmonization
Author: Robinson, Robert
ISNI:       0000 0004 9357 2584
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
In medical image analysis pipelines, Machine Learning (ML) is increasingly employed to process images and to derive clinically relevant biomarkers and quantitative measures. Results of these analyses are used to inform clinical decision-making from diagnosis to treatment planning. It is imperative that ML models work reliably, and their results are adequately scrutinized to avoid introducing bias or inaccuracy into downstream tasks. Image segmentation is a key component of many medical image analysis pipeline. For a segmentor to be reliable, it must be possible to identify instances when it has failed. This work presents an extensive validation of an approach for automated Quality Control (QC) of Cardiac Magnetic Resonance Imaging (CMR) segmentations using a reverse testing strategy, which is modified from Reverse Classification Accuracy (RCA) [164]. The method is evaluated on data from the UKBB imaging study, demonstrating its effectiveness at large scale in the absence of Ground Truth (GT). It achieves over 95% accuracy classifying segmentation quality using Dice Similarity Coefficient (DSC). It may be desirable to obtain an immediate assessment of segmentation quality in high throughput pipelines, in pharmaceutical trials, or in the clinic to inform operators when acquired images yield poor quality analysis results. This study presents a method for real-time automated QC using Deep Learning (DL)-based frameworks which directly predict DSCs and proxy RCA scores, achieving MAE = 0.03 and MAE = 0.14 respectively. A model may fail because it is evaluated on data drawn from a different distribution to that on which it was trained, as with scans acquired from different sites. This work introduces and evaluates an unpaired, unsupervised method for Domain Adaptation (DA) on multi-site neuroimaging data, aiming to mitigate the degradation in model performance due to Domain Shift (DS) caused by population and acquisition changes. Over 20% performance is recovered in a classification task when applying constrained appearance and spatial transformations with Image and Spatial Transformer Networks (ISTNs).
Supervisor: Glocker, Benjamin ; Rueckert, Daniel Sponsor: Engineering and Physical Sciences Research Council ; GlaxoSmithKline
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.815277  DOI:
Share: