Image segmentation

The Healthy Hearts Consortium dataset comprises over 9,000 CMR scans, which we analysed using pre-defined segmentation protocols with the intention of replicating current established clinical practice. CMR segmentation was performed at a core CMR lab (William Harvey Research Institute, NIHR Barts Biomedical Research Centre, QMUL). For the image analysis, we used a specific prototype (version 5.14 prototype) of a certified software (cvi42™, Circle Cardiovascular Imaging Inc., Calgary, AB, Canada) with batch processing capability.

Short axis images covering both left and right ventricle were analysed using three pre-defined segmentation protocols selected to capture established clinical practices (Figure 3), all of which are endorsed by the Society for Cardiovascular Magnetic Resonance:

  1. Smooth: Segmentation of endocardial contours, where papillary muscles and trabecular tissue are excluded from the LV mass (included in the blood pool).
  2. Papillary: Smooth endocardial contours with inclusion of papillary muscles (but not trabecular tissue) in the LV mass.
  3. Anatomical: Inclusion of both papillary muscles and trabecular tissue in the LV mass.

Figure 3. Illustration of segmentation methods

Figure 3 footnote: Illustration of ventricular segmentation protocols for short-axis CMR images. Smooth contours exclude papillary muscles and trabecular tissue from the ventricular mass, assigning them to the blood pool (Left). Papillary contours include papillary muscles within the myocardium but not the trabecular tissue (Middle). Anatomical contours incorporate both papillary muscles and trabecular tissue into the myocardial mass (Right). These segmentation approaches align with standards set by the Society for Cardiovascular Magnetic Resonance, demonstrating variability in myocardial mass and volume calculation based on contour definitions.

Why were three segmentation methods implemented?

The use of artificial intelligence (AI) in CMR post-processing has been on the rise, as more software packages offer automated or semi-automated solutions. However, the normal values generated by these software packages are usually based on manual segmentation, which can be time-consuming and prone to errors. Until recently, no AI solution was robust enough to be used for this purpose, and there have been concerns about the reproducibility of different segmentation protocols when segmented manually or using AI. Furthermore, there is lack of consensus in the CMR community regarding the optimal segmentation method, with several approaches currently used in clinical practice.

To address these critical issues, we re-segmented the images in a uniform way using the three clinically used segmentation protocols and we aimed to decrease technical sources of discrepancies and establish normal values for clinical use with quality checked AI based methods. Image analysis was with automated tools from two widely used post-processing software (CircleCVI, NEOSOFT), followed by visual quality control. The imaging data using NEOSOFT post-processing software will be made available in early spring. The AI-based segmentation method was quality checked by experts in the field to ensure the accuracy and reliability of the results. Our research provides readily accessible sex, age, and ethnicity-stratified normal values for LV, RV, LA, and RA volumes, mass, and function.

Quality control

A statistical quality control approach was implemented for the UK Biobank healthy subset, using a 3-standard deviation (3SD) threshold to remove extreme biologically implausible values. Manual visual quality control was performed for all other contributing cohorts by two expert readers using a custom-built Shiny app developed in-house. This app provides a platform for quality scoring across three domains: image acquisition, image planning, and segmentation. Each dimension is scored on a three-level score system where 1 = ‘perfect’, 2 = ‘satisfactory’, and 3 = ‘unacceptable for clinical use’. All values that received a quality score of 3, indicating poor quality, were removed from the analysis.