Deep Learning to Estimate Cardiac Magnetic Resonance-Derived Left Ventricular Mass

Shaan Khurshid, Sam Friedman, James P. Pirruccello, Paolo Di Achille, Nathaniel Diamant, Christopher D. Anderson, Patrick T. Ellinor, Puneet Batra, Jennifer E. Ho, Anthony Philippakis, Steven A. LubitzmedRxiv 2020 · DOI

Abstract

Background: Cardiac magnetic resonance (CMR) is the gold standard for left ventricular hypertrophy (LVH) diagnosis. CMR-derived LV mass can be estimated using proprietary algorithms (e.g., inlineVF), but their accuracy and availability may be limited. Objective: To develop an open-source deep learning model to estimate CMR-derived LV mass. Methods: Within participants of the UK Biobank prospective cohort undergoing CMR, we trained two convolutional neural networks to estimate LV mass. The first (ML4Hreg) performed regression informed by manually labeled LV mass, while the second (ML4Hseg) performed LV segmentation informed by inlineVF contours. We compared ML4Hreg, ML4Hseg, and inlineVF against manually labeled LV mass within an independent holdout set. Results: ML4Hseg reproduced manually labeled LV mass more accurately (r=0.864, MAE 10.41g) than ML4Hreg (r=0.843, MAE 10.51g) and inlineVF (r=0.795, MAE 14.30g). LVH defined using ML4Hseg demonstrated the strongest associations with hypertension, atrial fibrillation, and heart failure. Conclusions: ML4Hseg is an open-source deep learning model providing automated quantification of CMR-derived LV mass, facilitating broad cardiovascular discovery.

Introduction

Left Ventricular Hypertrophy (LVH)—the thickening of the heart's main pumping chamber—is a critical clinical indicator for cardiovascular events, including heart failure, stroke, and sudden cardiac death. While Cardiac Magnetic Resonance (CMR) imaging is considered the "gold standard" for measuring LV mass and diagnosing LVH, its widespread application in research and clinical practice is hindered by the labor-intensive nature of image analysis.

Traditionally, measuring LV mass requires a specialist to manually trace the heart's boundaries (segmentation) across multiple image slices. This process is too slow and expensive for large-scale population studies. While proprietary automated tools like Siemens' inlineVF exist, they often operate as "black boxes" with limited transparency, variable accuracy, and restricted accessibility. This study introduces an open-source, high-performing deep learning alternative designed to automate LV mass estimation at scale.

Methods

Researchers developed and compared two distinct deep learning architectures based on Convolutional Neural Networks (CNNs) using data from the UK Biobank (UKB):

  1. Direct Regression (ML4Hreg): This model was designed to output a single numerical value for LV mass directly from raw CMR images. It was trained on a high-quality "gold standard" subset of 5,065 individuals whose images were manually labeled by experts.
  2. Semantic Segmentation (ML4Hseg): Utilizing a U-Net style architecture, this model identifies every pixel belonging to the myocardium (heart muscle). It was trained on a larger set of 33,071 individuals using contours generated by the inlineVF algorithm. LV mass is then calculated by determining the volume of identified pixels and applying the known density of heart tissue (1.05 g/cm³).

The performance of both models, along with the proprietary inlineVF tool, was validated against an independent holdout set of 891 individuals with manual expert labels.

Key Results

Model Accuracy and Correlation

The segmentation-based model, ML4Hseg, emerged as the most accurate tool for LV mass estimation, outperforming both the regression model and the proprietary software:

  • Correlation: ML4Hseg achieved a Pearson correlation of r = 0.864 (95% CI 0.847-0.880), compared to r = 0.843 for ML4Hreg and r = 0.795 for inlineVF.
  • Error Rates: The Mean Absolute Error (MAE) for ML4Hseg was 10.41g, significantly lower than the 14.30g error observed with the proprietary inlineVF tool (p < 0.01).

Clinical Utility and Disease Association

Beyond technical accuracy, the researchers evaluated the clinical relevance of the automated measurements. LVH defined by the ML4Hseg model showed the strongest statistical associations with prevalent cardiovascular diseases in a logistic regression adjusted for age and sex:

  • Hypertension: Odds Ratio (OR) 2.76 (95% CI 2.51-3.04)
  • Atrial Fibrillation: OR 1.75 (95% CI 1.37-2.20)
  • Heart Failure: OR 4.53 (95% CI 3.16-6.33)

Discussion

Segmentation vs. Regression

A key finding of this study is that the segmentation model (ML4Hseg) outperformed the regression model (ML4Hreg), despite being trained on "noisier" labels from a proprietary algorithm rather than expert manual labels. This suggests that the spatial information inherent in segmentation tasks provides a more robust learning signal for deep learning models than simple scalar regression.

Scalability and Open Science

By automating the quantification process, the researchers generated LV mass estimates for 38,574 individuals in the UK Biobank—a task that would be practically impossible using manual methods. Furthermore, by making the model architecture and weights publicly available, the authors provide a transparent and accessible tool for the broader research community, overcoming the limitations of proprietary "black box" software.

Limitations

While the results are promising, the study notes that UK Biobank participants tend to be healthier than the general population, resulting in a lower distribution of LV mass than might be seen in clinical settings. Additionally, external validation is required to ensure the model generalizes across different MRI scanner manufacturers and more diverse patient demographics.

Conclusion

The ML4Hseg model provides a scalable, open-source solution for the automated estimation of LV mass from CMR images. By accurately identifying individuals at risk for hypertension and heart failure, such deep learning models can facilitate large-scale cardiovascular discovery and improve the efficiency of clinical research.

Summary updated

Abstract

ABSTRACT Background Cardiac magnetic resonance (CMR) is the gold standard for left ventricular hypertrophy (LVH) diagnosis. CMR-derived LV mass can be estimated using proprietary algorithms (e.g., inlineVF), but their accuracy and availability may be limited. Objective To develop an open-source deep learning model to estimate CMR-derived LV mass. Methods Within participants of the UK Biobank prospective cohort undergoing CMR, we trained two convolutional neural networks to estimate LV mass. The first (ML4H reg ) performed regression informed by manually labeled LV mass (available in 5,065 individuals), while the second (ML4H seg ) performed LV segmentation informed by inlineVF contours. We compared ML4H reg , ML4H seg , and inlineVF against manually labeled LV mass within an independent holdout set using Pearson correlation and mean absolute error (MAE). We assessed associations between CMR-derived LVH and prevalent cardiovascular disease using logistic regression adjusted for age and sex. Results We generated CMR-derived LV mass estimates within 38,574 individuals. Among 891 individuals in the holdout set, ML4H seg reproduced manually labeled LV mass more accurately (r=0.864, 95% CI 0.847-0.880; MAE 10.41g, 95% CI 9.82-10.99) than ML4H reg (r=0.843, 95% CI 0.823-0.861; MAE 10.51, 95% CI 9.86-11.15, p=0.01) and inlineVF (r=0.795, 95% CI 0.770-0.818; MAE 14.30, 95% CI 13.46-11.01, p&lt;0.01). LVH defined using ML4H seg demonstrated the strongest associations with hypertension (odds ratio 2.76, 95% CI 2.51-3.04), atrial fibrillation (1.75, 95% CI 1.37-2.20), and heart failure (4.53, 95% CI 3.16-6.33). Conclusions ML4H seg is an open-source deep learning model providing automated quantification of CMR-derived LV mass. Deep learning models characterizing cardiac structure may facilitate broad cardiovascular discovery.
BibTeX
@misc{khurshid2020deep,
  author = {Shaan Khurshid and Sam Friedman and James P. Pirruccello and Paolo Di Achille and Nathaniel Diamant and Christopher D. Anderson and Patrick T. Ellinor and Puneet Batra and Jennifer E. Ho and Anthony Philippakis and Steven A. Lubitz},
  title = {{Deep Learning to Estimate Cardiac Magnetic Resonance-Derived Left Ventricular Mass}},
  year = {2020},
  doi = {10.1101/2020.12.18.20248364},
  howpublished = {medRxiv},
}
Figure