Source-Resilient Joint Learning Framework for Preserving Stable Generalization on Diverse Ultrasonic Source Scenarios

Area of Science:

• Medical imaging
• Machine learning
• Computer vision

Background:

• Joint learning on diverse ultrasonic data faces challenges due to source heterogeneity and feature inconsistency, limiting generalization.
• Existing methods lack source-resilience, leading to performance degradation with varied ultrasound data.
• Limited single-source data variations and ultrasound imaging interference further hinder generalization.

Purpose of the Study:

• To propose a novel source-resilient joint learning framework to enhance generalization in diverse ultrasonic source scenarios.
• To address data heterogeneity, feature inconsistency, limited data variation, and ultrasound interference.
• To provide a generalizable framework for improving joint learning performance in medical imaging.

Main Methods:

• A three-stage framework: source transforming (1-to-N transformation), feature enhancement (manifold-constraint normalization module, task-consistent attention module, adaptive feature-shifting module), and ultrasound-hybrid linear mapping (speckle randomization, Monge-Kantorovitch linear mapping).
• Manifold-constraint normalization module (MCNM) minimizes manifold-based loss to address heterogeneity.
• Task-consistent attention module (TCAM) uses self-attention for multi-scale feature sharing to tackle inconsistency.
• Adaptive feature-shifting module (AFSM) provides feature-level augmentation for limited single-source data.
• Ultrasound-hybrid linear mapping (USmapping) randomizes ultrasonic style to mitigate imaging interference.

Main Results:

• The framework achieved superior performance on eight ultrasound datasets across multiple centers and scanners.
• Segmentation task: Weighted average Dice Similarity Coefficient (DSC_WAvg) of 75.7%.
• Classification task: Weighted average Area Under the Receiver Operating Characteristic curve (AUROC_WAvg) of 68.8%.

Conclusions:

• The proposed source-resilient joint learning framework effectively enhances generalization in diverse ultrasonic scenarios.
• The framework demonstrates significant improvements over previous methods in both segmentation and classification tasks.
• This approach holds potential as a general solution for robust joint learning in challenging medical imaging applications.