Prior-guided multi-expert consensus fusion for multi-center thyroid nodule classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Thyroid nodule classification in ultrasound imaging faces challenges due to variations in devices and protocols across different medical centers.
• These variations cause distribution shifts in data, impacting the performance of AI models.
• Developing robust models for cross-center classification is crucial for reliable thyroid cancer diagnosis.

Purpose of the Study:

• To develop a robust Prior-Guided Multi-Expert Consensus Fusion Network (PMCF) for accurate and generalizable thyroid nodule classification across multi-center ultrasound data.
• To address the challenges posed by entangled visual patterns and distribution shifts in cross-center datasets.
• To improve the stability and accuracy of diagnostic predictions in diverse clinical settings.

Main Methods:

• Proposed a Prior-Guided Multi-Expert Consensus Fusion Network (PMCF) incorporating a Prior-Guided Routing Mechanism and a Consensus Fusion Mechanism.
• The routing mechanism uses structure-aware spatial priors to modulate deep features across multiple expert branches, promoting feature disentanglement.
• The fusion mechanism aggregates expert outputs using inter-expert agreement and adaptive gating for enhanced robustness under domain shifts.

Main Results:

• The PMCF model demonstrated superior performance in classification accuracy and generalization compared to state-of-the-art methods on three multi-center thyroid ultrasound datasets.
• Ablation studies confirmed that prior-guided feature modulation enhances representation diversity.
• Consensus-based fusion was shown to improve prediction stability, validating the combined approach.

Conclusions:

• The proposed PMCF effectively addresses multi-center classification challenges in thyroid ultrasound imaging by combining structured priors with expert collaboration.
• The method offers improved robustness and generalization capabilities, paving the way for more reliable AI-assisted thyroid nodule diagnosis.
• The study highlights the significance of adaptive feature modulation and consensus fusion for cross-domain medical image analysis.