Towards Building a Trustworthy Deep Learning Framework for Medical Image Analysis.

Area of Science:

• Artificial Intelligence in Medicine
• Computer Vision
• Deep Learning for Medical Imaging

Background:

• Deep learning (DL) in medical AI offers diagnostic, predictive, and prognostic potential.
• Challenges in medical image analysis include limited/imbalanced data and the need for trustworthy models.
• Model performance and trust are critical for clinical adoption of AI.

Purpose of the Study:

• To introduce TRUDLMIA, a trustworthy deep learning framework for medical image analysis.
• To leverage self-supervised learning for feature extraction and a novel surrogate loss function for trust.
• To build trustworthy AI models with optimal performance for medical imaging tasks.

Main Methods:

• Developed TRUDLMIA, a framework integrating self-supervised learning and a novel surrogate loss function.
• Validated the framework on benchmark datasets for pneumonia, COVID-19, and melanoma detection.
• Conducted ablation studies, cross-validation, and result visualization to assess contributions.

Main Results:

• TRUDLMIA models demonstrated highly competitive performance, outperforming task-specific models.
• The framework significantly improved model performance (up to 21%) and model trust (up to 5%).
• Ablation studies confirmed the effectiveness of the proposed modules.

Conclusions:

• TRUDLMIA provides a robust solution for trustworthy deep learning in medical image analysis.
• The framework can advance AI applications in public health crises, improving diagnostics and patient outcomes.
• TRUDLMIA supports researchers and clinicians in enhancing healthcare quality through reliable AI.