DOMINO: Domain-aware loss for deep learning calibration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models achieve state-of-the-art performance in medical imaging.
• Model calibration is frequently overlooked, posing risks in high-stakes medical applications.
• Uncalibrated models provide unreliable confidence estimates, hindering clinical trust.

Purpose of the Study:

• To propose a novel domain-aware loss function for calibrating deep learning models in medical imaging.
• To enhance the reliability and trustworthiness of AI in healthcare.
• To reduce the potential for dangerous failures in automated diagnostic systems.

Main Methods:

• Development of a domain-aware loss function incorporating class similarity.
• Implementation of a class-wise penalty mechanism sensitive to inter-class relationships.
• Evaluation of the proposed loss function on medical imaging tasks.

Main Results:

• The proposed loss function significantly improves model calibration across medical imaging tasks.
• The approach reduces the likelihood of high-consequence errors by penalizing risky misclassifications.
• Domain-aware calibration leads to more reliable confidence scores from deep learning models.

Conclusions:

• Domain-aware loss functions are crucial for safe and effective deployment of deep learning in medicine.
• The proposed method offers a practical solution for enhancing the reliability of medical AI.
• Improved model calibration translates to safer clinical decision-making and better patient outcomes.