Cardiac Magnetic Resonance Imaging-Derived Phenotypes (CMR IDPs)

Introduction

UK Biobank’s imaging project is collecting multi-modality imaging data on over 100,000 participants, including cardiac magnetic resonance (CMR) scans. To make this more accessible for researchers, key measurements have been extracted from raw image data as imaging-derived phenotypes (IDPs). For information on all imaging modalities and associated IDPs, refer to the Imaging Data article. This article focusses specifically on the IDPs derived from CMR. It provides an overview on what CMRs are, what CMR IDPs are, where these IDPs can be found, and how they were derived.

What are CMRs?

CMR is a non-invasive imaging technique used to produce images of the heart. The UK Biobank’s CMR protocol is detailed in the paper by Petersen et al. (2016), designed to perform a range of imaging sequences to assess cardiac structure, function, blood flow, and myocardial tissue characterisation. Further details on the CMR scan procedure and protocol can be found in Resource 349 and Resource 15146. The raw DICOM images are available for each imaging sequence in fields 20207-20214 on Showcase. These are also listed under Category 1015, marked with a “‡”, to indicate bulk data.

CMR imaging sequences

(UK Biobank field names with alternative term in brackets where relevant)

• Aortic distensibility images
• Blood flow images (Blood flow above the aortic valve)
• Cine tagging images (Note: Acquisition stopped after 52,685 scans from 48,003 unique participants)
• Experimental shMOLLI sequence images (T1 mapping)
• Left ventricular outflow tract images
• Long-axis heart images (Cine LAX (long-axis views))
• Scout images for heart MRI (planning images)
• Short-axis heart images (Cine SAX (short-axis stack))

What are CMR IDPs and what do they mean?

CMR IDPs are quantitative measures extracted from CMR. These avoid the need for specialist processing by researchers who are not image scientists. Manual processing can often be time-consuming, labour-intensive, and subject to inter- and intra-observer variability. IDPs address these challenges by converting raw imaging files into quantitative measures that can be readily used in subsequent analyses. For example, instead of a raw MRI scan of the heart, an IDP could show the maximum volume of a participant’s left atrium in millilitres. These IDPs can be generated through automated algorithms applied to raw MRI scans, allowing researchers to analyse cardiac phenotypes at scale directly as data fields without performing their own image analysis.

UK Biobank’s CMR IDPs cover a range of cardiovascular features and can be broadly classified into six categories under Category 523:

1. Volume and size measurements
2. Wall thickness
3. Function
4. Strain metrics
5. Vascular measures
6. Tissue composition and mapping

Some IDP fields may appear to measure the same variable, but they present as separate fields as the method of derivation is different. For example, left atrial maximum volume appears under multiple field IDs (24110 and 31075) as they are derived using different pipelines.

Note: AHA refers to the American Heart Association’s 16-segment model, which divides the left ventricle into standardised regions using a short-axis view of the myocardial wall. Details on the segments can be found in the following: https://www.ahajournals.org/doi/10.1161/hc0402.102975. Longitudinal strain is calculated from the 4-chamber view motion tracking, using six segments: basal septal, basal lateral, mid septal, mid lateral, apical septal, and apical lateral, as defined in Bai et al.'s paper.

Where can IDPs be found?

CMR scans can be found in Category 102. CMR IDPs are primary located under the following categories:

• Category 157 (derived from a pipeline by Bai et al. (2020))
• Category 162 (derived from Petersen et al. (2017); Biasiolli et al. (2019); Ferdian et al. (2020); Bard et al. (2021); Aung et al. (2022))
• Category 133 (automatically derived using InlineVF software)

Each IDP has a unique field ID and description on Showcase, which researchers can select and work with on UK Biobank Research Analysis Platform (RAP).

How were these IDPs created?

The CMR IDPs are generated using automated pipelines developed to extract quantitative measures from the raw images. This section provides a brief overview of the methods used to derive the IDPs in categories 157, 162 and 133. Please refer to their specific papers for full details.

Category 157

Bai et al. (2020)

They developed an automated machine-learning pipeline to analyse UK Biobank CMR scans. Their pipeline was built upon a convolutional neural network (CNN) model for image segmentation, and image analysis was performed on short-axis, long-axis and aortic cine images for 26,893 participants. The network models were trained on manually annotated cardiac images. The data were originally in DICOM format and then converted to NIfTI for analysis in the pipeline. The pipeline was used to segment the left ventricle, right ventricle, left atrium, right atrium, ascending aorta, and descending aorta. From these contours, phenotypes such as volumes, ejection fraction, myocardial mass, myocardial wall thickness, and aortic cross-sectional areas are derived. As part of this analysis, the AHA 16-segment model is used to divide the left ventricle into standardised regions with a short-axis view of the myocardial wall. Longitudinal strain is calculated from the 4-chamber view motion tracking, using six segments: basal septal, basal lateral, mid septal, mid lateral, apical septal, and apical lateral, as defined in Bai et al.'s paper. For more details on the AHA model, please refer to this paper. Their pipeline provides 82 quantitative imaging phenotypes characterising the structure and function of the heart and aorta for each participant. The image analysis code is publicly available on their GitHub.

Category 162

Petersen et al. (2017)

They conducted a manual analysis of CMR scans in 804 healthy Caucasian UK Biobank participants to generate reference ranges for cardiac structure and function. All image analysis was performed using cvi42 software by Circle Cardiovascular Imaging Inc. and involved two core laboratories (London and Oxford, UK). Following a strict quality control protocol and consensus review for discrepancies, the team manually contoured the left ventricle, right ventricle, left atria, and right atria. The study also assessed intra- and inter-observer variability, and this was deemed “good to excellent”. They produced manually derived phenotypes for left and right ventricular end-diastolic volumes, ejection fractions, and atrial volumes.

Basiolli et al. (2019)

They developed an automated method to localise and segment the ascending and descending aorta in the transverse balanced Steady State Free Precession cine images and estimate aortic distensibility. This method was used on the first 5100 aortic cine scans from 4,996 UK Biobank participants, and incorporated built-in quality control by flagging low-quality scans.

Ferdian et al. (2020)

They used 4508 short-axis CMR-tagged images from 5,065 UK Biobank participants. Their method used two neural networks to estimate the left ventricular circumferential and radial strain. The framework involved a CNN and then a combination of a recurrent neural network and a CNN. This provided values for peak systolic circumferential strain in the apical, basal, and mid-ventricular regions of the left ventricle. Their code is publicly available on GitHub.

Bard et al. (2021)

They developed a neural network using a Multi-residual U-net (MultiResUNet) base architecture for automated segmentation of pericardial adipose tissue on four-chamber cine images. They also incorporated a monte carlo sampling to derive a measure of uncertainty that can serve as quality control. The study included 42,928 participants in their analysis and provide data for area of pericardial fat.

Aung et al. (2022)

They trained their data using Petersen et al.’s manual segmentation data from the short-axis cine images. 29,506 European UK Biobank participants who did not have pre-existing heart failure or myocardial infarction were included in their analysis. They used the segmentation pipeline developed by Bai et al. to derive right ventricle and right atrium volumes, stroke volume, and ejection fraction.

Category 133

InlineVF

The MRI scanner itself produced automated data on left ventricular size and function using Siemen’s syngo InlineVF software (Siemens Healthcare, Erlangen, Germany, version D13A) during cine acquisition. Derived data from this software include end diastolic and end systolic volumes, ejection fraction, and stroke volume. Please note that the inline VF results are fully automated without any expert quality control. Caution is advised to use results without any visual quality control or linear bias correction as recommended in the following paper by Suinesiaputra et al. (2018): Fully-automated left ventricular mass and volume MRI analysis in the UK Biobank population cohort: evaluation of initial results. 

If you need additional guidance or support please submit a ticket