DOI: 10.1161/circ.148.suppl_1.12928 ISSN: 0009-7322

Abstract 12928: Deep Learning-Derived Characterization of Lung Tissue Predicts Future Diagnosis of Heart Failure

Cody R Hou, Joel Ramo, Shinwan Kany, Samuel Friedman, Patrick T Ellinor, James Pirruccello
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: Cardiovascular and respiratory diseases are intrinsically coupled. Scalable methods for quantifying cardiac and pulmonary structures may improve our understanding of cardiopulmonary disease.

Aims: We sought to quantify textural parameters of the lung fields in cardiovascular magnetic resonance imaging (cMRI) from the UK Biobank. We then sought to determine their genetic basis and their relationship with heart failure.

Methods: We fine-tuned a convolutional neural network to segment left and right lung on cMRIs performed during expiratory breath hold. A random sample of 300 coronal slices from unique individuals was manually annotated for fine-tuning. The model was applied to 912,320 coronal cMRI images. Segmentation masks were stacked, and lung volume and texture metrics were quantified from the resulting 3D masks using the pyradiomics Python package.

Results: After quality control, we computed lung traits in 45,179 participants. Mean lung volume was 2.19 ± 0.49 L in women and 2.87 ± 0.68 L in men. After adjustment for age and sex, lung volume was associated with forced vital capacity (FVC) (β = +0.30 L per SD, p < 1.1 х 10 -308 ). After adjustment for age and sex, body size traits (e.g. standing height and heel bone mineral density) were associated with increased lung volume, whereas adiposity traits (e.g. body mass index and visceral adipose tissue) were associated with decreased lung volume. Greater heterogeneous and coarse lung textures were associated with an increased risk of heart failure (1-SD hazard ratios: 1.21 (95% CI [1.12, 1.31]) and 1.32 (95% CI [1.18, 1.47]), respectively). This association was still significant among non-smokers and after adjustment for left ventricular ejection fraction.

Conclusions: Deep-learning derived measures of lung texture are linked to future heart failure. Our findings suggest that heart failure may manifest as changes in lung tissue that are detectable on MRI before changes in cardiac function.

More from our Archive