Dynamic Ensemble Transfer Learning with Multi-view Ultrasonography for Improving Thyroid Cancer Diagnostic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Diagnostic decision-making integrates facts and clinician experience.
• Diverse clinical experience in multi-disciplinary models mitigates knowledge gaps.
• Current computer-aided diagnostic systems struggle with diverse datasets, limiting reliability.

Purpose of the Study:

• Propose a dynamic ensemble transfer learning system to enhance diagnostic decision-making reliability.
• Simulate data and knowledge diversity within the system's training and structure.
• Improve diagnostic accuracy for thyroid cancer, a rapidly rising malignancy.

Main Methods:

• Developed a system with self-directed model selection, dynamic weighting, and unified weighted ensemble averaging.
• Pre-trained individual networks using two multi-view thyroid ultrasonography datasets from over 700 cross-national patients.
• Evaluated the fine-tuned ensemble model on external thyroid nodule data and assessed generalization through data resampling.

Main Results:

• The proposed ensemble model achieved promising performance with an area under the curve (AUC) between 0.87 and 0.93.
• Benchmarking demonstrated superior performance compared to existing studies.
• The system improved diagnostic reliability in thyroid cancer care.

Conclusions:

• The dynamic ensemble transfer learning system effectively simulates data diversity for improved diagnostic reliability.
• The approach shows significant potential in enhancing thyroid cancer diagnosis and guiding management.
• This method offers a robust solution for computer-aided diagnostic systems facing data heterogeneity.