Using Machine Learning to Identify Noninvasive Motion-Based Biomarkers of Cardiac Function
The aim of this project is to apply state-of-the-art imaging, motion analysis and machine learning techniques to characterise the motion of the heart as it beats. A population-based, spatio-temporal atlas of cardiac motion has been developed that enables subtle variations in cardiac motion to be identified and quantified. This analysis forms the basis for the development of novel noninvasive biomarkers for the stratification of patients with cardiovascular disease. A notable success to date has been the development of biomarkers that can predict response to cardiac resynchronisation therapy with a success rate of 91% (Peressutti et al, Medical Image Analysis), whilst the current clinical selection criteria result in around one third of patients not responding positively to the treatment.
- M. Sinclair, D. Peressutti, E. Puyol-Antón, W. Bai, S. Rivolo, J. Webb, S. Claridge, T. Jackson, D. Nordsletten, M. Hadjicharalambous, E. Kerfoot, C. A. Rinaldi, D. Rueckert, A. P. King, "Myocardial Strain Computed at Multiple Spatial Scales from Tagged Magnetic Resonance Imaging: Estimating Cardiac Biomarkers for CRT Patients", Medical Image Analysis 43:169-185, 2017. (open access paper)
- D. Peressutti, M. Sinclair, W. Bai, T. Jackson, J. Ruijsink, D. Nordsletten, L. Asner, M. Hadjicharalambous, C. A. Rinaldi, D. Rueckert, A. P. King, "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 35:669-684, 2017. (open access paper)
- M. Sinclair, D. Peressutti, E. Puyol-Antón, W. Bai, D. Nordsletten, M. Hadjicharalambous, E. Kerfoot, T. Jackson, S. Claridge, C. A. Rinaldi, D. Rueckert, and A. P. King, "Learning Optimal Spatial Scales for Cardiac Strain Analysis Using a Motion Atlas", Proceedings MICCAI STACOM, 2016.
- D. Peressutti, W. Bai, T. Jackson, M. Sohal, C. A. Rinaldi, D. Rueckert, and A. P. King, "Prospective Identification of CRT Super Responders Using a Motion Atlas and Random Projection Ensemble Learning", Proceedings MICCAI, 2015. (paper, presentation)
- D. Peressutti, W. Bai, W. Shi, C. Tobon-Gomez, T. Jackson, M. Sohal, A. Rinaldi, D. Rueckert, and A. King, "Towards Left Ventricular Scar Localisation Using Local Motion Descriptors", Proceedings MICCAI STACOM, 2015. (paper)
More recent work has focused on the incorporation of multi-modal data into the atlas, in the form of MR and US imaging (Puyol-Anton et al, Medical Image Analysis). Multiview dimensionality reduction methods such as Canonical Correlation Analysis and Partial Least Squares Regression were utilised to embed high-dimensional data from MR and US into a common low-dimensional space where they could be directly compared. This work was subsequently extended to exploit multiview learning techniques to directly diagnose disease from multimodal data sources (Puyol-Anton et al, IEEE Transactions on Biomedical Engineering). A novel regional multiview learning algorithm was developed that was based upon analsis of cardiac cycle motion within the 17 segments of the AHA model.
- E. Puyol-Antón, B. Ruijsink, B. Gerber, M. S. Amzulescu, H. Langet, M. De Craene, J. A. Schnabel, P. Piro, A. P. King, "Regional Multi-view Learning for Cardiac Motion Analysis: Application to Identification of Dilated Cardiomyopathy Patients", IEEE Transactions on Biomedical Engineering, 2018. (paper)
- E. Puyol-Antón, M. Sinclair, B. Gerber, M. S. Amzulescu, H. Langet, M. De Craene, P. Aljabar, J. A. Schnabel, P. Piro, A. P. King, "Multiview Machine Learning Using an Atlas of Cardiac Cycle Motion", Proceedings MICCAI STACOM, pp3-11, 2018. (paper)
- E. Puyol-Antón, M. Sinclair, B. Gerber, M. S. Amzulescu, H. Langet, M. De Craene, P. Aljabar, P. Piro, A. P. King, "A Multimodal Spatiotemporal Cardiac Motion Atlas from MR and Ultrasound Data", Medical Image Analysis 40:96-110, 2017. (open access paper)
- E. Puyol-Antón, D. Peressutti, P. Aljabar, M. De Craene, P. Piro and A. P. King, "Towards a Multimodal Cardiac Motion Atlas for Biomarker Extraction", Proceedings ISBI, 2016. (paper)
Automated Quality Control of Cine MR Data
(Funded by EPSRC grant EP/P001009/1)
Recently large-scale databases of imaging and non-imaging data have started to become more widely available. Such databases have great potential for learning more about clinical risk factors and disease processes. The UK Biobank will eventually consist of data from 100,000 volunteers and contains cine and tagged cardiac MR data as well as a wide range of other personal data. However, the size of this database means that it will inevitably contain poor quality data, which could impact upon any analysis performed upon the data unless the poor quality data can be excluded. Manual exclusion is tedious, time-consuming and subject to error. In this work we investigate the use of machine learning techniques for performing automated quality control of cine MR data.
The first technique aims to detect poorly planned 4-chamber cine MR images. One common feature of such images is the presence of the left ventricular outflow tract (LVOT) in the image. We trained a deep convolutional neural network (CNN) on a training set of images featuring good/poor planning. The resulting classifier achieved an accuracy of approximately 83%.
- I. Oksuz, B. Ruijsink, E. Puyol-Antón, M. Sinclair, D. Rueckert, J. A. Schnabel, A. P. King, "Automatic Left Ventricular Outflow Tract Classification for Accurate Cardiac MR Planning", Proceedings ISBI, pp462-465, 2018. (paper)
Subsequent work focused on the presence of motion-related artefacts in the cine MR images. Such artefacts can be caused by mistriggering due to a faulty ECG signal, or by patient breathing during the acquisition. They typically appear as blurring artefacts in the images. Again, UK Biobank data were used to train a CNN classifier to identify these artefacts. However, there is high class imbalance with this problem - there are many more good quality images than artefact images. Class imbalance in training data is known to lead to bias in the performance of the resulting classifier. To address this problem we developed a novel technique for augmenting the training data using synthetically generated artefact images. These images were generated in k-space using a similar process to that which causes the artefacts - we then used a curriculum learning strategy to gradually teach the model about progressively more difficult classification examples.
- I. Oksuz, B. Ruijsink, E. Puyol-Antón, J. R. Clough, G. Cruz, A. Bustin, C. Prieto, R. Botnar, D. Rueckert, J. A. Schnabel, A. P. King, "Automatic CNN-based detection of cardiac MR motion artefacts using k-space data augmentation and curriculum learning", Medical Image Analysis, 55:136-147, 2019. (paper)
- I. Oksuz, B. Ruijsink, E. Puyol-Antón, A. Bustin, G. Cruz, C. Prieto, D. Rueckert, J. A. Schnabel, A. P. King, "Deep Learning Using K-space Based Data Augmentation for Automated Cardiac MR Motion Artefact Detection", Proceedings MICCAI, pp250-258, 2018. (paper)
Automatic Quantification of Cardiac Function
As an example application of the automated quality control techniques discussed above, we exploited the clinical information contained in the UK Biobank database to investigate factors associated with cardiac health. This involved forming a motion atlas from around 6000 subjects and statistically analysing the links between atlas-derived motion descriptors and the clinical information. Associations were found with body fat percentage, basal metabolic rate, hypertension, smoking status and alcohol intake frequency, and these associations were not identifiable by analysis of the conventional metric of ejection fraction.
- E. Puyol-Antón, B. Ruijsink, H. Langet, M. De Craene, P. Piro, J. A. Schnabel, A. P. King, "Learning Associations Between Clinical Information and Motion-based Descriptors Using a Large Scale MR-derived Cardiac Motion Atlas", Proceedings MICCAI STACOM, pp94-102, 2018. (paper)
Subsequently, the power of deep learning was used to construct a fully-automated, quality-controlled pipeline for quantification of a wide range of cardiac biomarkers. Prior to this exciting work, automatic quantification of cardiac function was limited to mostly global metrics such as volumes and ejection fraction, which do not provide a detailed description of cardiac health. We applied deep learning-based automated segmentation and a range of other machine learning tools to robustly estimate a richer range of biomarkers such as peak ejection rate, peak filling rate, atrial contribution and strains from the UK Biobank cohort. For this first time, healthy reference ranges were computed for these biomarkers.
- B. Ruijsink, E. Puyol-Antón, I. Oksuz, M. Sinclair, W. Bai, J. A. Schnabel, R. Razavi, A. P. King, "Fully Automated, Quality-Controlled Cardiac Analysis From CMR: Validation and Large-Scale Application to Characterize Cardiac Function", JACC: Cardiovascular Imaging, 2019 (in press). (open access paper)
- B. Ruijsink, E. Puyol-Antón, Y. Li, W. Bai, E. Kerfoot, R. Razavi, A. P. King, "Quality-aware Semi-supervised Learning for CMR Segmentation", Proceedings MICCAI STACOM, 2020. (Arxiv paper)
We next developed a similar pipeline for automated quantification of ShMOLLI T1 data. This pipeline also featured robust quality control steps, this time based upon segmentation uncertainty. Disease-specific reference ranges from the UK Biobank dataset were published for the first time.
- E. Puyol-Antón, B. Ruijsink, C. F. Baumgartner, P.-G. Masci, M. Sinclair, E. Konukoglu, R. Razavi, A. P. King, "Automated Quantification of Myocardial Tissue Characteristics From Native T1 Mapping Using Neural Networks With Uncertainty-based Quality-control", Journal of Cardiovascular Magnetic Resonance, 22:60, 2020. (open access paper)
Interpretable Machine Learning in Cardiology
Although deep learning techniques have produced some impressive results in tasks such as disease classification, one of their main drawbacks is the lack of interpretability of their outputs. Put simply, they can make decisions but often cannot explain in human-understandable terms how those decisions were made. In this work we addressed this limitation. We showed how a disease classification network could be built using a variational autoencoder, and that the latent space of the autoencoder could be analysed to link the decision with other high-level concepts. In experiments on the UK Biobank cardiac cine MR data, we showed that biomarkers relating to ventricular ejection and filling rates had a large contribution to the task of classifying coronary artery disease.
- J. R. Clough, I. Oksuz, E. Puyol-Antón, B. Ruijsink, A. P. King, J. A. Schnabel, "Global and Local Interpretability for Cardiac MRI Classification", Proceedings MICCAI, 2019. (Arxiv paper)
In related work, cardiac functional biomarkers, which were automatically estimated using the deep learning based approach from Ruijsink et al, were used in combination with a variational autoencoder network. The network's loss function was defined to also perform a regression against systolic blood pressure in its latent space. This novel architecture enabled a detailed investigation of the links between systolic blood pressure and the functional biomarkers.
- E. Puyol-Antón, B. Ruijsink, J. R. Clough, I. Oksuz, D. Rueckert, R. Razavi, A. P. King, "Assessing the Impact of Blood Pressure on Cardiac Function Using Interpretable Biomarkers and Variational Autoencoders", Proceedings MICCAI STACOM, 2019. (Arxiv paper)