Medical Imaging

EPSRC Centre for Doctoral Training


301 - Explaining the predictions of deep learning models in cardiology

1st Supervisor: Andy King
2nd Supervisor: Ben Glocker

Project Aim:

Our intention is to produce a computer aided diagnosis tool to assist clinicians in diagnosing heart disease and planning its treatment using motion and/or deformation information. The tool will be based on the latest deep learning techniques, and the application of deep learning to cardiac motion analysis represents the first main novelty of this project. The second novelty lies in the explanatory power of the tool. By offering clear and intuitive information that can help to generate explanations for its outputs it is hoped that one of the most significant obstacles to the clinical translation of machine learning techniques in medicine (i.e. trust) can be overcome.

Project Background:

The use of machine learning techniques in cardiology is an active research area and recently state-of-the-art results have been achieved by applying the latest ‘deep learning’ techniques to tasks such as segmentation of the myocardium and blood pool [1]. Machine learning has also been used to identify and exploit the useful information carried by the motion and deformation of the heart as it beats [2,3]. The intuition behind these techniques is that certain pathologies will have an influence on the electrical activation of the heart, and so observing the mechanical motion and deformation that results from this activation can offer insights into the underlying pathology.

Applications demonstrated to date include characterising septal flash [2] and predicting response to cardiac resynchronisation therapy [3]. To date, these techniques have been mostly based on more traditional machine learning methods and deep learning has not been extensively investigated to exploit motion information in cardiology.

Whilst machine learning techniques have produced impressive results, a significant problem remains. Often, the techniques that produce the most accurate results lack one important feature that is important for the clinical acceptance of new technology: explanatory power. Put simply, most machine learning classifiers or regressors are able to make predictions but are not able to explain in human-interpretable terms how the prediction was arrived at. The reason for this is that often the input to such algorithms are very high-dimensional data, and optimal performance is achieved by using complex, often non-linear, hierarchical combinations of subsets of these data. This reasoning process is opaque, and clinicians are reluctant to base clinical decisions upon recommendations from such a black-box machine.

[1] Litjens et al, A Survey on Deep Learning in Medical Image Analysis, CoRR, 2017.
[2] Duchateau et al, A spatiotemporal statistical atlas of motion for the quantification of abnormal myocardial tissue velocities, Medical Image Analysis, 2011.
[3] Peressutti et al, A Framework for Combining a Motion Atlas with Non-Motion Information to Learn Clinically Useful Biomarkers: Application to Cardiac Resynchronisation Therapy Response Prediction, Medical Image Analysis, 2017.

Image and video hosting by TinyPic

Figure: An overview of our vision of a system for explaining the predictions made by deep learning techniques in cardiology.