Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable Representation Learning of Cardiac MRI via Attribute Regularization

Maxime Di Folco, Cosmin I. Bercea, Emily Chan, Julia A. Schnabel

TL;DR

Interpreting latent representations in medical imaging is crucial for clinician trust, yet VAEs often produce blurry reconstructions. This work introduces AR-SIVAE, which integrates an attribute regularization loss $\mathcal{L}_{attr}$ into the Soft Introspective VAE framework, with $\alpha=2$, to align latent dimensions with clinically relevant attributes while maintaining sharp generation. Evaluated on UK Biobank cardiac MRI, AR-SIVAE achieves non-blurry reconstructions and improved interpretability of the latent space compared to baselines such as $\beta$-VAE, SIVAE, and Attri-VAE, evidenced by interpretable latent traversals and higher SCC metrics. The approach enables attribute-controlled latent factors suitable for downstream cardiac MRI tasks, though it relies on many hyperparameters and could benefit from extending to non morphometric attributes and downstream classification applications.

Abstract

Interpretability is essential in medical imaging to ensure that clinicians can comprehend and trust artificial intelligence models. Several approaches have been recently considered to encode attributes in the latent space to enhance its interpretability. Notably, attribute regularization aims to encode a set of attributes along the dimensions of a latent representation. However, this approach is based on Variational AutoEncoder and suffers from blurry reconstruction. In this paper, we propose an Attributed-regularized Soft Introspective Variational Autoencoder that combines attribute regularization of the latent space within the framework of an adversarially trained variational autoencoder. We demonstrate on short-axis cardiac Magnetic Resonance images of the UK Biobank the ability of the proposed method to address blurry reconstruction issues of variational autoencoder methods while preserving the latent space interpretability.

Interpretable Representation Learning of Cardiac MRI via Attribute Regularization

TL;DR

Interpreting latent representations in medical imaging is crucial for clinician trust, yet VAEs often produce blurry reconstructions. This work introduces AR-SIVAE, which integrates an attribute regularization loss into the Soft Introspective VAE framework, with , to align latent dimensions with clinically relevant attributes while maintaining sharp generation. Evaluated on UK Biobank cardiac MRI, AR-SIVAE achieves non-blurry reconstructions and improved interpretability of the latent space compared to baselines such as -VAE, SIVAE, and Attri-VAE, evidenced by interpretable latent traversals and higher SCC metrics. The approach enables attribute-controlled latent factors suitable for downstream cardiac MRI tasks, though it relies on many hyperparameters and could benefit from extending to non morphometric attributes and downstream classification applications.

Abstract

Interpretability is essential in medical imaging to ensure that clinicians can comprehend and trust artificial intelligence models. Several approaches have been recently considered to encode attributes in the latent space to enhance its interpretability. Notably, attribute regularization aims to encode a set of attributes along the dimensions of a latent representation. However, this approach is based on Variational AutoEncoder and suffers from blurry reconstruction. In this paper, we propose an Attributed-regularized Soft Introspective Variational Autoencoder that combines attribute regularization of the latent space within the framework of an adversarially trained variational autoencoder. We demonstrate on short-axis cardiac Magnetic Resonance images of the UK Biobank the ability of the proposed method to address blurry reconstruction issues of variational autoencoder methods while preserving the latent space interpretability.
Paper Structure (14 sections, 6 equations, 4 figures, 2 tables)

This paper contains 14 sections, 6 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Preliminaries:
  4. Attributed regularized VAE
  5. Soft Introspective Variational Autoencoder (SIVAE)
  6. Proposed method: AR-SIVAE
  7. Dataset and Implementation details
  8. Dataset:
  9. Attributes:
  10. Implementation details:
  11. Experiments and results
  12. Reconstruction performance
  13. Interpretability of the latent space
  14. Conclusion

Figures (4)

  • Figure 1: AR-SIVAE: Attribute regularized Soft Introspective Variational AutoEncoder (AR-SIVAE) combines attribute regularization within the SIVAE framework (an adversarially trained encoder-decoder network) to enhance the interpretability of the latent space while being able to generate non-blurry samples.
  • Figure 2: Illustration of the AR-SIVAE framework. Attri. Reg: Attribute regularization.
  • Figure 3: Qualitative evaluation of the reconstruction of two samples at and for $\beta$-VAE, , Attri-VAE and the proposed method AR-SIVAE. The first column corresponds to the ground-truth (GT). The SIVAE-based methods ($4^{th}$ and $5^{th}$ columns) overcome the blurry reconstruction of VAE-based methods ($2^{nd}$ and $3^{rd}$ columns).
  • Figure 4: Walk in the regularized latent dimensions of LV end-diastolic volume (first row) and RV end-systolic volume (second row) for Attri-VAE (left) and AR-SIVAE (right).