Revisiting the impact of sex imbalances on biased deep learning models in medical imaging

This repository contains the source code to reproduce the figures presented in the paper "Revisiting the impact of sex imbalances on biased deep learning models in medical imaging".

Instructions to reproduce results

Here, we describe how to reproduce the figures presented in the paper.

0. Run threshold_calc.ipynb.

1. Run fairness_calculation.ipynb. This notebook will create two tables, final_bias_favor_data.csv and final_overall_bias_favor_data.csv, and save them in the tables folder. The final_overall_bias_favor_data.csv is Table 1 from the paper.

2. Run fairness_correction.ipynb. This notebook will create the five tables listed below and save them in the tables folder. The table, train100%_female_images.csv, can contains the results to quickly reproduce Table 2 from the paper.

   1. train100%_female_images.csv
   2. train75%_female_images.csv
   3. train50%_female_images.csv
   4. train25%_female_images.csv
   5. train0%_female_images.csv

3. Run generate_plots.ipynb. This notebook will create the figures final_all_conditions_graph.pdf and final_sub_conditions_graph.pdf. The final_sub_conditions_graph.pdf figure is Fig. 2 from the paper.