Reinforced Computer-Aided Framework for Diagnosing Thyroid Cancer.

Area of Science:

• Endocrinology
• Medical Imaging
• Artificial Intelligence

Background:

• Thyroid cancer is a prevalent endocrine disease requiring early detection.
• Ultrasound examination is the primary method for early thyroid nodule screening.
• Current deep learning models for single ultrasound images lack accuracy and generalization.

Purpose of the Study:

• To propose a practical, diagnosis-oriented computer-aided diagnosis (CAD) framework for thyroid nodules.
• To enhance the accuracy and generalization of thyroid nodule classification using a novel approach.
• To develop a scalable framework capable of integrating diverse diagnostic information.

Main Methods:

• A framework combining collaborative deep learning and reinforcement learning was developed.
• Deep learning models were trained collaboratively on multiparty, privacy-preserving data.
• A reinforcement learning agent fused classification results, modeling diagnostic information as a Markov decision process (MDP).
• A dataset of two thousand thyroid ultrasound images was utilized for collaborative training.

Main Results:

• The proposed framework demonstrated promising performance in simulated experiments.
• Collaborative learning on large-scale, privacy-preserving medical data enhanced model robustness and generalization.
• The MDP-based fusion by the reinforcement learning agent achieved precise final diagnosis results.

Conclusions:

• The novel CAD framework effectively improves thyroid nodule diagnosis accuracy and generalization.
• The integration of collaborative learning and reinforcement learning offers a robust approach for medical image analysis.
• The framework's scalability allows for the incorporation of more diagnostic data sources for improved precision.