Challenges in Reducing Bias Using Post-Processing Fairness for Breast Cancer Stage Classification with Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast cancer is a leading global health concern for women.
• Deep learning models show promise in breast cancer diagnosis and treatment.
• Ensuring fairness in AI models across diverse populations is a significant challenge, particularly with underrepresented demographic groups in training data.

Purpose of the Study:

• To quantify bias in deep learning models used for breast cancer staging.
• To evaluate the effectiveness of post-processing calibration in mitigating identified biases.

Main Methods:

• Trained deep learning models to predict breast cancer stage using a dataset of 1000 biopsies from 842 patients.
• The dataset predominantly comprised data from White patients (over 70%).
• Assessed model performance disparities between White and non-White patient groups before and after calibration.

Main Results:

• Prior to calibration, all models exhibited superior performance for White patients compared to non-White patients.
• Post-processing calibration led to varied improvements, with only some models showing enhanced equitable performance.
• The study highlights inherent biases in AI models trained on non-diverse datasets.

Conclusions:

• Deep learning models for breast cancer staging can perpetuate racial bias due to data disparities.
• Post-processing calibration is not a universally effective solution for achieving fairness in medical AI.
• Addressing data diversity is crucial for developing equitable AI tools in breast cancer care.