Interpretable and backpropagation-free Green Learning for efficient multi-task echocardiographic segmentation and classification

TL;DR

This work tackles the need for accurate and trustworthy echocardiographic analysis by removing backpropagation from a multi-task framework that segments the left ventricle and classifies LVEF. It introduces MTGL, a Green Learning pipeline built around an unsupervised VoxelHop encoder, a coarse-to-fine residual segmentation decoder, and an XGBoost-based classifier, achieving state-of-the-art results on EchoNet-Dynamic (ACC $=0.943$, DSC $=0.912$) with only ~1.13M parameters. The method preserves interpretability through linear, voxel-level features and explicit energy-based feature selection, enabling inspection of how edge and motion cues drive segmentation and functional assessment. Practically, MTGL demonstrates high performance, edge-device friendliness, and reduced data and energy demands, supporting safer, faster, and more deployable cardiac AI in clinical workflows.

Abstract

Echocardiography is a cornerstone for managing heart failure (HF), with Left Ventricular Ejection Fraction (LVEF) being a critical metric for guiding therapy. However, manual LVEF assessment suffers from high inter-observer variability, while existing Deep Learning (DL) models are often computationally intensive and data-hungry "black boxes" that impede clinical trust and adoption. Here, we propose a backpropagation-free multi-task Green Learning (MTGL) framework that performs simultaneous Left Ventricle (LV) segmentation and LVEF classification. Our framework integrates an unsupervised VoxelHop encoder for hierarchical spatio-temporal feature extraction with a multi-level regression decoder and an XG-Boost classifier. On the EchoNet-Dynamic dataset, our MTGL model achieves state-of-the-art classification and segmentation performance, attaining a classification accuracy of 94.3% and a Dice Similarity Coefficient (DSC) of 0.912, significantly outperforming several advanced 3D DL models. Crucially, our model achieves this with over an order of magnitude fewer parameters, demonstrating exceptional computational efficiency. This work demonstrates that the GL paradigm can deliver highly accurate, efficient, and interpretable solutions for complex medical image analysis, paving the way for more sustainable and trustworthy artificial intelligence in clinical practice.

Figures (5)

• Figure 1: Data Distribution of classification task. Class 1 denotes as LVEF $>$ 50%, class 2 denotes as LVEF between 40–50% and class 3 denotes as LVEF between $<$ 40%.
• Figure 2: Multi-task Green Learning (MTGL) Model Architecture.
• Figure 3: LV segmentation accuracy DSC (a) and IoU (b) across models. Boxplots show the median, interquartile range (IQR), and overall range for 3D V-Net, 3D UNETR, 3D U-Net, 3D nnU-Net, and the proposed MTGL model. Our method yields the highest median DSC and IoU with the lowest variability, indicating superior and more consistent performance. Asterisks denote statistical significance versus our model (two-sided paired t-test, Holm–Bonferroni corrected); **** denotes $p<10^{-4}$.
• Figure 4: Interpretable multi-hop feature responses: Response of the most influential AC filter in the 1st and 4th hops of the VoxelHop encoder. The color bar visualizes the AC filter response computed on z-score standardized voxel intensities (arbitrary units, a.u.). Red indicates positive filter responses (stronger activation) and blue indicates negative responses (suppression), with the magnitude given by the color bar.
• Figure 5: The energy plot as a function of the number of AC filters. All numbers in this plot are computed from the saved spectra of the trained encoder, by concatenating per–node AC eigenvalues within each hop after DC removal and then applying the cumulative–energy rule.