A reinforcement learning model for AI-based decision support in skin cancer

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Machine Learning

Background:

• Diagnostic artificial intelligence (AI) decision support systems aim to improve clinical accuracy.
• Integrating human preferences into AI algorithms is an underexplored area for enhancing diagnostic capabilities.

Purpose of the Study:

• To investigate the potential of incorporating human preferences into AI decision support for medical diagnosis.
• To evaluate the effectiveness of reinforcement learning (RL) with human-derived reward structures in improving diagnostic AI performance for skin cancer detection.

Main Methods:

• Utilized reinforcement learning (RL) with nonuniform rewards and penalties based on expert-generated tables to balance diagnostic errors.
• Compared RL models against traditional supervised learning (SL) for skin cancer diagnosis.
• Assessed the impact of the RL model on dermatologist diagnostic accuracy and management decisions.

Main Results:

• RL significantly improved sensitivity for melanoma (61.4% to 79.5%) and basal cell carcinoma (79.4% to 87.1%).
• AI overconfidence was reduced while maintaining accuracy.
• Dermatologist correct diagnosis rates increased by 12.0%, and optimal management decisions improved from 57.4% to 65.3%.

Conclusions:

• Incorporating human preferences via RL can significantly enhance diagnostic AI performance in medical imaging.
• This approach improves diagnostic accuracy and supports better clinical decision-making by healthcare professionals.
• Findings suggest a promising direction for developing more effective and human-aligned AI diagnostic tools.