Cross-domain dynamic routing decoders for multi-domain generalization in ultrasound imaging

The heterogeneous quality of medical ultrasound (US) imaging across different geographical regions presents a significant challenge for developing robust artificial intelligence (AI) systems in healthcare. While high-income regions benefit from standardized imaging protocols and advanced equipment, resource-constrained environments often produce data with pronounced quality variations, limiting the generalization capabilities of conventional deep learning models. We introduce USHydraNet, a novel multi-decoder framework designed to manage the heterogeneity of multi-regional US images in medical image analysis.

USHydraNet integrates a vision transformer (ViT) encoder or UNet encoder with multiple decoders, optimizing feature extraction across diverse data distributions. The framework employs a dynamic routing paradigm that selects the optimal decoder output by analyzing image-level and feature-level statistical moments. This adaptive mechanism enables robust performance across varying data quality levels without compromising accuracy.

Experimental validation across four public and one private medical datasets demonstrates that the USHydraNet model achieves superior performance over baseline architectures. In classification tasks, Ablation studies revealed that all metrics of the USHydraNet model showed improvements ranging from 10% to 20%. Comparative evaluations indicated that the USHydraNet model outperformed the four other models in terms of metric gains. For segmentation tasks, ablation experiments revealed that USHydraNet improved Dice scores by 12.23% and 19.52%, and intersection over union (IoU) by 0.59% and 1.05% across two datasets. Comparative experiments demonstrated that the USHydraNet model outperformed the four other models in all metric improvements. Even on unfamiliar datasets, it maintained robust performance with 89.26% Dice and 84.14% IoU.

USHydraNet is a promising framework for reducing performance disparities in medical image analysis across regions with varying healthcare infrastructures. thereby promoting equitable access to AI-assisted diagnosis in resource-limited settings.