Multi-Scale Self-Supervised Consistency Training for Trustworthy Medical Imaging Classification.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Medical Imaging
• Deep Learning for Healthcare

Background:

• Neural networks show human-level performance in medical diagnosis.
• Miscalibration in neural networks leads to inaccurate confidence estimation, hindering clinical trust.
• Reliable AI tools are crucial for advancing medical diagnosis.

Purpose of the Study:

• To propose a novel neural network calibration framework to address miscalibration in medical AI.
• To enhance the trustworthiness and reliability of neural network predictions in medical contexts.
• To improve both calibration and classification performance of medical AI models.

Main Methods:

• Utilized multi-scale input images for neural network training.
• Integrated self-supervised consistency enforcement during the training process.
• Developed a general-purpose calibration framework applicable to various imaging modalities.

Main Results:

• Demonstrated significant enhancement of neural network calibration.
• Achieved concomitant improvements in model classification performance.
• Cultivated more robust feature spaces, enhancing model generalizability.
• Showcased the framework's applicability across diverse medical imaging types.

Conclusions:

• The proposed framework effectively improves neural network calibration and trustworthiness in medical AI.
• The method enhances classification performance and feature robustness.
• This general-purpose solution can be combined with other techniques for further refinement, advancing reliable AI in medicine.