Clinically Aware Learning: Ordinal Loss Improves Medical Image Classifiers.

Area of Science:

• Medical Imaging AI
• Machine Learning for Healthcare
• Breast Cancer Screening

Background:

• Breast Imaging Reporting and Data System (BI-RADS) is crucial for breast cancer detection.
• Current models often treat BI-RADS as nominal, ignoring its ordinal nature and the clinical impact of misclassification.
• The mismatch between model optimization and clinical severity is an underexplored issue.

Purpose of the Study:

• To evaluate if ordinal-aware loss functions enhance BI-RADS classification performance.
• To compare ordinal losses against standard cross-entropy under controlled conditions.
• To analyze the impact of dataset and label balancing on performance.

Main Methods:

• Systematic evaluation of ordinal-aware loss functions (e.g., Earth Mover Distance) versus cross-entropy.
• Unified training pipeline across multiple datasets with fixed architecture.
• Analysis of dataset and label balancing strategies.
• Performance measured by Area Under the Receiver Operating Characteristic Curve (AUROC) and macro-F1 scores.

Main Results:

• Balanced sampling during training significantly improved performance.
• Ordinal loss functions consistently outperformed conventional cross-entropy.
• Improvements were particularly noted in reducing severe misclassifications, enhancing clinical relevance.
• Earth Mover Distance (EMD) demonstrated superior performance.

Conclusions:

• Aligning learning objectives with the ordinal BI-RADS structure substantially improves classification accuracy.
• Ordinal-aware approaches enhance AI model robustness and clinical relevance in breast cancer screening.
• Emphasizes the importance of loss function design, regularization, and data balancing in medical AI.