Cardiac aging synthesis from cross-sectional data with conditional generative adversarial networks

Víctor M. Campello, Tian Xia, Xiao Liu, Pedro Sanchez, Carlos Martín-Isla, Steffen E. Petersen, Santi Seguí, Sotirios A. Tsaftaris, Karim LekadirFrontiers in Cardiovascular Medicine 2022 · DOI

Abstract

Age is a primary risk factor for cardiovascular disease. Understanding how the heart ages is crucial for intervention, but longitudinal data—tracking the same individuals over decades—is difficult and expensive to collect. This study proposes a conditional generative adversarial network (cGAN) that synthesizes older and younger versions of cardiac MRI scans using only cross-sectional data. Trained on over 14,000 scans from the UK Biobank, the model learns to induce anatomical changes consistent with clinical literature, such as thickening of the interventricular septum and enlargement of the aorta. We demonstrate that these synthetic images can be used to correct age bias in medical AI datasets, improving the performance of age-prediction models.

Introduction

Understanding how the human body ages is a critical step toward developing interventions for age-related conditions. In cardiac health, age is the primary risk factor for cardiovascular disease. While longitudinal studies—tracking the same individuals over many years—provide the most accurate data on aging, they are prohibitively expensive and slow to conduct. This research presents a novel approach: using a conditional generative adversarial network (cGAN) to synthesize the aging process of the heart using only cross-sectional data.

Methodology: The cGAN Framework

The authors propose a cGAN architecture designed for image-to-image translation, capable of "aging" or "rejuvenating" a single cardiac MRI scan.

Overview of the cGAN framework showing the Generator producing aged/rejuvenated images and the Discriminator evaluating realism.

Conditional Biasing and Architecture

The generator is built on a U-Net foundation, incorporating residual blocks and attention mechanisms. To precisely control the degree of aging, the researchers implemented a "conditional biasing" technique.

Detailed architecture of the generator block, illustrating the injection of age-gap information via sinusoidal embeddings and channel-wise addition.

As shown in the figure above, the target age gap is transformed into a vector using Transformer sinusoidal embeddings—a method typically used in Natural Language Processing to encode positional information. This vector is then injected into every layer of the network through channel-wise addition, allowing the model to modulate anatomical features based on the desired age shift.

Training and Stability

To maintain the unique anatomical "identity" of the patient's heart during transformation, the model utilizes a cycle-consistency loss. This forces the network to be able to reverse any aging or rejuvenation it performs, returning the image to its original state. Training is further stabilized using a Wasserstein GAN with Gradient Penalty (WGAN-GP), which helps the model produce realistic 3D morphological changes.

Data and Validation

The model was trained on 14,788 four-chamber view cardiac MRI scans from the UK Biobank, covering a demographic range of 45 to 82 years. To validate the results, the researchers compared the synthetic images against clinical literature and used a subset of 764 scans with manual segmentations to measure volumetric changes.

Key Findings

  • Clinical Consistency: Without explicit supervision, the model learned to modify the interventricular septum (thickening) and the aorta (enlargement), aligning perfectly with known clinical markers of cardiac aging.
  • High Predictive Accuracy: When synthetic "aged" images were processed by a pre-trained age regressor, the predicted ages were highly accurate, with a Mean Absolute Error (MAE) of approximately 5.0 years for a 10-year gap.
  • Dataset De-biasing: A significant practical application is the correction of age bias in AI training sets. By augmenting a young-biased dataset (90% young) with 25% synthetic "elderly" images, the researchers reduced age prediction error from 14.5 years to 11.0 years.

Limitations and Future Directions

Despite its success, the model currently operates on 2D slices, whereas cardiac aging is inherently a 3D process. Future iterations will aim for volumetric consistency across the entire heart. Additionally, while the synthetic images are highly realistic, the model occasionally produces minor artifacts or "hallucinations" in complex anatomical regions like the atria.

Conclusion

This work demonstrates that deep generative models can successfully bridge the gap between cross-sectional snapshots and longitudinal insights. By synthesizing realistic aging trajectories, this approach provides a powerful tool for both clinical research and the development of more robust, un-biased medical AI.

Summary updated

Abstract

Age has important implications for health, and understanding how age manifests in the human body is the first step for a potential intervention. This becomes especially important for cardiac health, since age is the main risk factor for development of cardiovascular disease. Data-driven modeling of age progression has been conducted successfully in diverse applications such as face or brain aging. While longitudinal data is the preferred option for training deep learning models, collecting such a dataset is usually very costly, especially in medical imaging. In this work, a conditional generative adversarial network is proposed to synthesize older and younger versions of a heart scan by using only cross-sectional data. We train our model with more than 14,000 different scans from the UK Biobank. The induced modifications focused mainly on the interventricular septum and the aorta, which is consistent with the existing literature in cardiac aging. We evaluate the results by measuring image quality, the mean absolute error for predicted age using a pre-trained regressor, and demonstrate the application of synthetic data for counter-balancing biased datasets. The results suggest that the proposed approach is able to model realistic changes in the heart using only cross-sectional data and that these data can be used to correct age bias in a dataset.
BibTeX
@inproceedings{campello2022cardiac,
  author = {Víctor M. Campello and Tian Xia and Xiao Liu and Pedro Sanchez and Carlos Martín-Isla and Steffen E. Petersen and Santi Seguí and Sotirios A. Tsaftaris and Karim Lekadir},
  title = {{Cardiac aging synthesis from cross-sectional data with conditional generative adversarial networks}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2022},
  doi = {10.3389/fcvm.2022.983091},
}
Figure