Esther Puyol-Anton

Tareen Dawood, Chen Chen, Baldeep S. Sidhua et al.

Quantifying uncertainty of predictions has been identified as one way to develop more trustworthy artificial intelligence (AI) models beyond conventional reporting of performance metrics. When considering their role in a clinical decision support setting, AI classification models should ideally avoid confident wrong predictions and maximise the confidence of correct predictions. Models that do this are said to be well-calibrated with regard to confidence. However, relatively little attention has been paid to how to improve calibration when training these models, i.e., to make the training strategy uncertainty-aware. In this work we evaluate three novel uncertainty-aware training strategies comparing against two state-of-the-art approaches. We analyse performance on two different clinical applications: cardiac resynchronisation therapy (CRT) response prediction and coronary artery disease (CAD) diagnosis from cardiac magnetic resonance (CMR) images. The best-performing model in terms of both classification accuracy and the most common calibration measure, expected calibration error (ECE) was the Confidence Weight method, a novel approach that weights the loss of samples to explicitly penalise confident incorrect predictions. The method reduced the ECE by 17% for CRT response prediction and by 22% for CAD diagnosis when compared to a baseline classifier in which no uncertainty-aware strategy was included. In both applications, as well as reducing the ECE there was a slight increase in accuracy from 69% to 70% and 70% to 72% for CRT response prediction and CAD diagnosis respectively. However, our analysis showed a lack of consistency in terms of optimal models when using different calibration measures. This indicates the need for careful consideration of performance metrics when training and selecting models for complex high-risk applications in healthcare.

Tiarna Lee, Esther Puyol-Anton, Bram Ruijsink et al.

In computer vision there has been significant research interest in assessing potential demographic bias in deep learning models. One of the main causes of such bias is imbalance in the training data. In medical imaging, where the potential impact of bias is arguably much greater, there has been less interest. In medical imaging pipelines, segmentation of structures of interest plays an important role in estimating clinical biomarkers that are subsequently used to inform patient management. Convolutional neural networks (CNNs) are starting to be used to automate this process. We present the first systematic study of the impact of training set imbalance on race and sex bias in CNN-based segmentation. We focus on segmentation of the structures of the heart from short axis cine cardiac magnetic resonance images, and train multiple CNN segmentation models with different levels of race/sex imbalance. We find no significant bias in the sex experiment but significant bias in two separate race experiments, highlighting the need to consider adequate representation of different demographic groups in health datasets.

Germain Morilhat, Naomi Kifle, Sandra FinesilverSmith et al.

In many low-to-middle income (LMIC) countries, ultrasound is used for assessment of pleural effusion. Typically, the extent of the effusion is manually measured by a sonographer, leading to significant intra-/inter-observer variability. In this work, we investigate the use of deep learning (DL) to automate the process of pleural effusion segmentation from ultrasound images. On two datasets acquired in a LMIC setting, we achieve median Dice Similarity Coefficients (DSCs) of 0.82 and 0.74 respectively using the nnU-net DL model. We also investigate the use of coordinate convolutions in the DL model and find that this results in a statistically significant improvement in the median DSC on the first dataset to 0.85, with no significant change on the second dataset. This work showcases, for the first time, the potential of DL in automating the process of effusion assessment from ultrasound in LMIC settings where there is often a lack of experienced radiologists to perform such tasks.

Zhen Yuan, Esther Puyol-Anton, Haran Jogeesvaran et al.

Abnormal spleen enlargement (splenomegaly) is regarded as a clinical indicator for a range of conditions, including liver disease, cancer and blood diseases. While spleen length measured from ultrasound images is a commonly used surrogate for spleen size, spleen volume remains the gold standard metric for assessing splenomegaly and the severity of related clinical conditions. Computed tomography is the main imaging modality for measuring spleen volume, but it is less accessible in areas where there is a high prevalence of splenomegaly (e.g., the Global South). Our objective was to enable automated spleen volume measurement from 2D cross-sectional segmentations, which can be obtained from ultrasound imaging. In this study, we describe a variational autoencoder-based framework to measure spleen volume from single- or dual-view 2D spleen segmentations. We propose and evaluate three volume estimation methods within this framework. We also demonstrate how 95% confidence intervals of volume estimates can be produced to make our method more clinically useful. Our best model achieved mean relative volume accuracies of 86.62% and 92.58% for single- and dual-view segmentations, respectively, surpassing the performance of the clinical standard approach of linear regression using manual measurements and a comparative deep learning-based 2D-3D reconstruction-based approach. The proposed spleen volume estimation framework can be integrated into standard clinical workflows which currently use 2D ultrasound images to measure spleen length. To the best of our knowledge, this is the first work to achieve direct 3D spleen volume estimation from 2D spleen segmentations.

Tiarna Lee, Esther Puyol-Anton, Bram Ruijsink et al.

Artificial intelligence (AI) methods are being used increasingly for the automated segmentation of cine cardiac magnetic resonance (CMR) imaging. However, these methods have been shown to be subject to race bias, i.e. they exhibit different levels of performance for different races depending on the (im)balance of the data used to train the AI model. In this paper we investigate the source of this bias, seeking to understand its root cause(s) so that it can be effectively mitigated. We perform a series of classification and segmentation experiments on short-axis cine CMR images acquired from Black and White subjects from the UK Biobank and apply AI interpretability methods to understand the results. In the classification experiments, we found that race can be predicted with high accuracy from the images alone, but less accurately from ground truth segmentations, suggesting that the distributional shift between races, which is often the cause of AI bias, is mostly image-based rather than segmentation-based. The interpretability methods showed that most attention in the classification models was focused on non-heart regions, such as subcutaneous fat. Cropping the images tightly around the heart reduced classification accuracy to around chance level. Similarly, race can be predicted from the latent representations of a biased segmentation model, suggesting that race information is encoded in the model. Cropping images tightly around the heart reduced but did not eliminate segmentation bias. We also investigate the influence of possible confounders on the bias observed.

Devran Ugurlu, Esther Puyol-Anton, Bram Ruijsink et al.

Domain shift refers to the difference in the data distribution of two datasets, normally between the training set and the test set for machine learning algorithms. Domain shift is a serious problem for generalization of machine learning models and it is well-established that a domain shift between the training and test sets may cause a drastic drop in the model's performance. In medical imaging, there can be many sources of domain shift such as different scanners or scan protocols, different pathologies in the patient population, anatomical differences in the patient population (e.g. men vs women) etc. Therefore, in order to train models that have good generalization performance, it is important to be aware of the domain shift problem, its potential causes and to devise ways to address it. In this paper, we study the effect of domain shift on left and right ventricle blood pool segmentation in short axis cardiac MR images. Our dataset contains short axis images from 4 different MR scanners and 3 different pathology groups. The training is performed with nnUNet. The results show that scanner differences cause a greater drop in performance compared to changing the pathology group, and that the impact of domain shift is greater on right ventricle segmentation compared to left ventricle segmentation. Increasing the number of training subjects increased cross-scanner performance more than in-scanner performance at small training set sizes, but this difference in improvement decreased with larger training set sizes. Training models using data from multiple scanners improved cross-domain performance.

Ines Machado, Esther Puyol-Anton, Kerstin Hammernik et al.

Cine cardiac MRI is routinely acquired for the assessment of cardiac health, but the imaging process is slow and typically requires several breath-holds to acquire sufficient k-space profiles to ensure good image quality. Several undersampling-based reconstruction techniques have been proposed during the last decades to speed up cine cardiac MRI acquisition. However, the undersampling factor is commonly fixed to conservative values before acquisition to ensure diagnostic image quality, potentially leading to unnecessarily long scan times. In this paper, we propose an end-to-end quality-aware cine short-axis cardiac MRI framework that combines image acquisition and reconstruction with downstream tasks such as segmentation, volume curve analysis and estimation of cardiac functional parameters. The goal is to reduce scan time by acquiring only a fraction of k-space data to enable the reconstruction of images that can pass quality control checks and produce reliable estimates of cardiac functional parameters. The framework consists of a deep learning model for the reconstruction of 2D+t cardiac cine MRI images from undersampled data, an image quality-control step to detect good quality reconstructions, followed by a deep learning model for bi-ventricular segmentation, a quality-control step to detect good quality segmentations and automated calculation of cardiac functional parameters. To demonstrate the feasibility of the proposed approach, we perform simulations using a cohort of selected participants from the UK Biobank (n=270), 200 healthy subjects and 70 patients with cardiomyopathies. Our results show that we can produce quality-controlled images in a scan time reduced from 12 to 4 seconds per slice, enabling reliable estimates of cardiac functional parameters such as ejection fraction within 5% mean absolute error.

Esther Puyol-Anton, Bram Ruijsink, Stefan K. Piechnik et al.

The subject of "fairness" in artificial intelligence (AI) refers to assessing AI algorithms for potential bias based on demographic characteristics such as race and gender, and the development of algorithms to address this bias. Most applications to date have been in computer vision, although some work in healthcare has started to emerge. The use of deep learning (DL) in cardiac MR segmentation has led to impressive results in recent years, and such techniques are starting to be translated into clinical practice. However, no work has yet investigated the fairness of such models. In this work, we perform such an analysis for racial/gender groups, focusing on the problem of training data imbalance, using a nnU-Net model trained and evaluated on cine short axis cardiac MR data from the UK Biobank dataset, consisting of 5,903 subjects from 6 different racial groups. We find statistically significant differences in Dice performance between different racial groups. To reduce the racial bias, we investigated three strategies: (1) stratified batch sampling, in which batch sampling is stratified to ensure balance between racial groups; (2) fair meta-learning for segmentation, in which a DL classifier is trained to classify race and jointly optimized with the segmentation model; and (3) protected group models, in which a different segmentation model is trained for each racial group. We also compared the results to the scenario where we have a perfectly balanced database. To assess fairness we used the standard deviation (SD) and skewed error ratio (SER) of the average Dice values. Our results demonstrate that the racial bias results from the use of imbalanced training data, and that all proposed bias mitigation strategies improved fairness, with the best SD and SER resulting from the use of protected group models.

Zhen Yuan, Esther Puyol-Anton, Haran Jogeesvaran et al.

Sickle Cell Disease (SCD) is one of the most common genetic diseases in the world. Splenomegaly (abnormal enlargement of the spleen) is frequent among children with SCD. If left untreated, splenomegaly can be life-threatening. The current workflow to measure spleen size includes palpation, possibly followed by manual length measurement in 2D ultrasound imaging. However, this manual measurement is dependent on operator expertise and is subject to intra- and inter-observer variability. We investigate the use of deep learning to perform automatic estimation of spleen length from ultrasound images. We investigate two types of approach, one segmentation-based and one based on direct length estimation, and compare the results against measurements made by human experts. Our best model (segmentation-based) achieved a percentage length error of 7.42%, which is approaching the level of inter-observer variability (5.47%-6.34%). To the best of our knowledge, this is the first attempt to measure spleen size in a fully automated way from ultrasound images.

Bram Ruijsink, Esther Puyol-Anton, Ye Li et al.

One of the challenges in developing deep learning algorithms for medical image segmentation is the scarcity of annotated training data. To overcome this limitation, data augmentation and semi-supervised learning (SSL) methods have been developed. However, these methods have limited effectiveness as they either exploit the existing data set only (data augmentation) or risk negative impact by adding poor training examples (SSL). Segmentations are rarely the final product of medical image analysis - they are typically used in downstream tasks to infer higher-order patterns to evaluate diseases. Clinicians take into account a wealth of prior knowledge on biophysics and physiology when evaluating image analysis results. We have used these clinical assessments in previous works to create robust quality-control (QC) classifiers for automated cardiac magnetic resonance (CMR) analysis. In this paper, we propose a novel scheme that uses QC of the downstream task to identify high quality outputs of CMR segmentation networks, that are subsequently utilised for further network training. In essence, this provides quality-aware augmentation of training data in a variant of SSL for segmentation networks (semiQCSeg). We evaluate our approach in two CMR segmentation tasks (aortic and short axis cardiac volume segmentation) using UK Biobank data and two commonly used network architectures (U-net and a Fully Convolutional Network) and compare against supervised and SSL strategies. We show that semiQCSeg improves training of the segmentation networks. It decreases the need for labelled data, while outperforming the other methods in terms of Dice and clinical metrics. SemiQCSeg can be an efficient approach for training segmentation networks for medical image data when labelled datasets are scarce.

James R. Clough, Ilkay Oksuz, Esther Puyol-Anton et al.

Deep learning methods for classifying medical images have demonstrated impressive accuracy in a wide range of tasks but often these models are hard to interpret, limiting their applicability in clinical practice. In this work we introduce a convolutional neural network model for identifying disease in temporal sequences of cardiac MR segmentations which is interpretable in terms of clinically familiar measurements. The model is based around a variational autoencoder, reducing the input into a low-dimensional latent space in which classification occurs. We then use the recently developed `concept activation vector' technique to associate concepts which are diagnostically meaningful (eg. clinical biomarkers such as `low left-ventricular ejection fraction') to certain vectors in the latent space. These concepts are then qualitatively inspected by observing the change in the image domain resulting from interpolations in the latent space in the direction of these vectors. As a result, when the model classifies images it is also capable of providing naturally interpretable concepts relevant to that classification and demonstrating the meaning of those concepts in the image domain. Our approach is demonstrated on the UK Biobank cardiac MRI dataset where we detect the presence of coronary artery disease.

lkay Oksuz, James Clough, Bram Ruijsink et al.

In fully sampled cardiac MR (CMR) acquisitions, motion can lead to corruption of k-space lines, which can result in artefacts in the reconstructed images. In this paper, we propose a method to automatically detect and correct motion-related artefacts in CMR acquisitions during reconstruction from k-space data. Our correction method is inspired by work on undersampled CMR reconstruction, and uses deep learning to optimize a data-consistency term for under-sampled k-space reconstruction. Our main methodological contribution is the addition of a detection network to classify motion-corrupted k-space lines to convert the problem of artefact correction to a problem of reconstruction using the data consistency term. We train our network to automatically correct for motion-related artefacts using synthetically corrupted cine CMR k-space data as well as uncorrupted CMR images. Using a test set of 50 2D+time cine CMR datasets from the UK Biobank, we achieve good image quality in the presence of synthetic motion artefacts. We quantitatively compare our method with a variety of techniques for recovering good image quality and showcase better performance compared to state of the art denoising techniques with a PSNR of 37.1. Moreover, we show that our method preserves the quality of uncorrupted images and therefore can be also utilized as a general image reconstruction algorithm.

Ilkay Oksuz, Bram Ruijsink, Esther Puyol-Anton et al.

Good quality of medical images is a prerequisite for the success of subsequent image analysis pipelines. Quality assessment of medical images is therefore an essential activity and for large population studies such as the UK Biobank (UKBB), manual identification of artefacts such as those caused by unanticipated motion is tedious and time-consuming. Therefore, there is an urgent need for automatic image quality assessment techniques. In this paper, we propose a method to automatically detect the presence of motion-related artefacts in cardiac magnetic resonance (CMR) cine images. We compare two deep learning architectures to classify poor quality CMR images: 1) 3D spatio-temporal Convolutional Neural Networks (3D-CNN), 2) Long-term Recurrent Convolutional Network (LRCN). Though in real clinical setup motion artefacts are common, high-quality imaging of UKBB, which comprises cross-sectional population data of volunteers who do not necessarily have health problems creates a highly imbalanced classification problem. Due to the high number of good quality images compared to the relatively low number of images with motion artefacts, we propose a novel data augmentation scheme based on synthetic artefact creation in k-space. We also investigate a learning approach using a predetermined curriculum based on synthetic artefact severity. We evaluate our pipeline on a subset of the UK Biobank data set consisting of 3510 CMR images. The LRCN architecture outperformed the 3D-CNN architecture and was able to detect 2D+time short axis images with motion artefacts in less than 1ms with high recall. We compare our approach to a range of state-of-the-art quality assessment methods. The novel data augmentation and curriculum learning approaches both improved classification performance achieving overall area under the ROC curve of 0.89.

Ilkay Oksuz, Bram Ruijsink, Esther Puyol-Anton et al.

Quality assessment of medical images is essential for complete automation of image processing pipelines. For large population studies such as the UK Biobank, artefacts such as those caused by heart motion are problematic and manual identification is tedious and time-consuming. Therefore, there is an urgent need for automatic image quality assessment techniques. In this paper, we propose a method to automatically detect the presence of motion-related artefacts in cardiac magnetic resonance (CMR) images. As this is a highly imbalanced classification problem (due to the high number of good quality images compared to the low number of images with motion artefacts), we propose a novel k-space based training data augmentation approach in order to address this problem. Our method is based on 3D spatio-temporal Convolutional Neural Networks, and is able to detect 2D+time short axis images with motion artefacts in less than 1ms. We test our algorithm on a subset of the UK Biobank dataset consisting of 3465 CMR images and achieve not only high accuracy in detection of motion artefacts, but also high precision and recall. We compare our approach to a range of state-of-the-art quality assessment methods.

Esther Puyol-Anton, Bram Ruijsink, Helene Langet et al.

The availability of large scale databases containing imaging and non-imaging data, such as the UK Biobank, represents an opportunity to improve our understanding of healthy and diseased bodily function. Cardiac motion atlases provide a space of reference in which the motion fields of a cohort of subjects can be directly compared. In this work, a cardiac motion atlas is built from cine MR data from the UK Biobank (~ 6000 subjects). Two automated quality control strategies are proposed to reject subjects with insufficient image quality. Based on the atlas, three dimensionality reduction algorithms are evaluated to learn data-driven cardiac motion descriptors, and statistical methods used to study the association between these descriptors and non-imaging data. Results show a positive correlation between the atlas motion descriptors and body fat percentage, basal metabolic rate, hypertension, smoking status and alcohol intake frequency. The proposed method outperforms the ability to identify changes in cardiac function due to these known cardiovascular risk factors compared to ejection fraction, the most commonly used descriptor of cardiac function. In conclusion, this work represents a framework for further investigation of the factors influencing cardiac health.