A Novel Distant Domain Transfer Learning Framework for Thyroid Image Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodule diagnosis relies on ultrasound imaging, but interpretation varies among radiologists.
• Deep learning offers potential but faces challenges with limited annotated medical data.
• Traditional transfer learning struggles with significant domain differences, leading to negative transfer.

Purpose of the Study:

• To propose a novel transfer learning method, distant domain high-level feature fusion (DHFF), for improved thyroid nodule classification.
• To address the limitations of annotated data acquisition and negative transfer in medical image analysis.
• To enhance the accuracy and reliability of automated thyroid nodule diagnosis.

Main Methods:

• Developed a novel DHFF model for transfer learning in medical image analysis.
• Implemented a technique to reduce distribution distance between source and target domains while preserving domain characteristics.
• Validated the DHFF model using multiple public and private thyroid ultrasound datasets.

Main Results:

• The DHFF model achieved a classification accuracy of up to 88.92% for thyroid nodule diagnosis.
• Demonstrated an improvement of up to 8% compared to existing transfer and distant transfer algorithms.
• Successfully leveraged auxiliary source domains to enhance target domain performance.

Conclusions:

• The proposed DHFF method effectively overcomes challenges in medical image transfer learning.
• DHFF offers a robust solution for improving the accuracy of computer-aided diagnosis systems for thyroid nodules.
• This approach facilitates more reliable and consistent thyroid nodule classification, reducing observer variability.