CoReEcho: Continuous Representation Learning for 2D+time Echocardiography Analysis

TL;DR

CoReEcho tackles EF regression from 2D+time echocardiograms by enforcing continuous representations in a two-stage training framework. A representation-aware loss encourages the encoder to organize embeddings by EF, while a lightweight regressor maps these embeddings to EF, with a second stage refining the regression head after freezing the encoder. The approach achieves state-of-the-art EF regression on EchoNet-Dynamic (MAE $=3.90$, $R^2=82.44$, RMSE $=5.13$) and demonstrates robust transferability to CAMUS and HMC-QU, highlighting strong out-of-domain generalization. By producing explainable, continuous embeddings, CoReEcho supports potential echocardiography foundation-model use and facilitates distribution-shift detection and downstream adaptation in clinical settings.

Abstract

Deep learning (DL) models have been advancing automatic medical image analysis on various modalities, including echocardiography, by offering a comprehensive end-to-end training pipeline. This approach enables DL models to regress ejection fraction (EF) directly from 2D+time echocardiograms, resulting in superior performance. However, the end-to-end training pipeline makes the learned representations less explainable. The representations may also fail to capture the continuous relation among echocardiogram clips, indicating the existence of spurious correlations, which can negatively affect the generalization. To mitigate this issue, we propose CoReEcho, a novel training framework emphasizing continuous representations tailored for direct EF regression. Our extensive experiments demonstrate that CoReEcho: 1) outperforms the current state-of-the-art (SOTA) on the largest echocardiography dataset (EchoNet-Dynamic) with MAE of 3.90 & R2 of 82.44, and 2) provides robust and generalizable features that transfer more effectively in related downstream tasks. The code is publicly available at https://github.com/fadamsyah/CoReEcho.

Figures (5)

• Figure 1: UMAP of learned feature embeddings on the EchoNet ouyang2020video test set. Unlike EchoCoTr, CoReEcho assembles continuous representations in the embedding space with respect to true EF, thus mitigating clinically irrelevant relations.
• Figure 2: CoReEcho consists of two training stages: 1) Optimize $\mathcal{F}_E$ to learn continuous relation between echocardiogram clips while simultaneously teaching $\mathcal{F}_R$ to estimate EF. 2) Perform MLP probing to further optimize $\mathcal{F}_R$ while $\mathcal{F}_E$ is frozen.
• Figure 2: Grad-CAM samples on the EchoNet-Dynamic test set. CoReEcho can place a higher emphasis on the LV region. CoReEcho also exhibits reduced focus on backgrounds, in contrast to EchoCoTr.
• Figure 3: Grad-CAMgrad-cam. CoReEcho$\mathcal{F}_E$ exhibits reduced focus on backgrounds and more focus on the LV. We provide additional samples in Supplementary Materials.
• Figure 4: Grad-CAM visualization on the HMC-QU dataset using the MLP probing setting.