Interpretable and backpropagation-free Green Learning for efficient multi-task echocardiographic segmentation and classification
Jyun-Ping Kao, Jiaxin Yang, C.-C. Jay Kuo, Jonghye WooPurpose
This study develops an interpretable, backpropagation-free multi-task Green Learning (MTGL) framework for automated echocardiographic analysis. It addresses the need for reliable left ventricular (LV) segmentation and left ventricular ejection fraction (LVEF) classification while reducing the parameter burden, energy demand and opacity associated with conventional deep learning (DL) models. The objective is to determine whether a unified feed-forward framework can deliver clinically relevant accuracy, computational efficiency and interpretable feature representations.
Design/methodology/approach
The framework uses a four-hop, PCA/Saab-based unsupervised VoxelHop encoder followed by a coarse-to-fine XGBoost residual-regression decoder for LV segmentation and a pooled multi-scale XGBoost classifier for three-class LVEF classification. Experiments used the predefined EchoNet-Dynamic training, validation and test partitions. MTGL was evaluated against matched 3D V-Net, 3D U-Net, 3D UNETR and 3D nnU-Net baselines using classification accuracy and balanced accuracy, and Dice similarity coefficient (DSC) and intersection over union (IoU).
Findings
On the EchoNet-Dynamic test set, MTGL achieved 94.3% accuracy for three-class LVEF classification, and 0.912 DSC for LV segmentation, outperforming all compared baselines. It used 1.13 million parameters, versus 10.47–49.08 million for the DL comparators, and showed the lowest measured training energy use and carbon emissions. Hop ablation further showed complementary multi-scale information, with the full feature set yielding the highest classification accuracy.
Originality/value
Unlike end-to-end backpropagation-based DL methods, MTGL combines an unsupervised linear VoxelHop representation with task-specific XGBoost decoders in a single multi-task echocardiographic pipeline. The PCA/Saab direct current and alternating current feature decomposition enables direct inspection of learned spatial and spatiotemporal signals, linking shallow features to anatomical boundaries and deeper features to global contractile dynamics. The study therefore demonstrates that a compact, interpretable Green Learning system can jointly support LV segmentation and LVEF grouping with strong in-domain performance and lower measured training environmental cost.