Deep reinforcement learning framework for thoracic diseases classification via prior knowledge guidance.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Thoracic Disease Diagnosis

Background:

• Chest X-rays are crucial for diagnosing thoracic diseases.
• Limited labeled data hinders accurate automated diagnosis.
• Existing methods struggle with data scarcity and generalization.

Purpose of the Study:

• To develop a novel deep reinforcement learning (DRL) framework for thorax disease diagnosis.
• To enhance diagnostic accuracy by incorporating prior knowledge and continuous learning.
• To address the challenge of limited labeled data in medical imaging.

Main Methods:

• A DRL framework guiding diagnostic agents with prior knowledge.
• Acquiring prior knowledge from pre-trained or similar domain models.
• Enabling continuous model updates mimicking human learning.
• Utilizing reinforcement learning for agent exploration.

Main Results:

• Reduced dependence on target domain data through knowledge transfer.
• Improved diagnostic accuracy via continuous exploration.
• Enhanced model generalization capability with limited data.
• Competitive performance on NIH ChestX-ray 14 and CheXpert datasets.

Conclusions:

• The proposed DRL framework effectively diagnoses thoracic diseases.
• Prior knowledge integration and continuous learning improve accuracy and generalization.
• The method shows promise for clinical application, especially with data limitations.