Catalog of Regulatory Science Tools to Help Assess New Medical Devices
This regulatory science tool presents two methods to amplify Artificial Intelligence (AI) / Machine Learning (ML) model bias to enable the evaluation of bias mitigation methods on models with varying amounts of performance bias.
Technical Description
bias.myti.report is a python-based software tool that assists with the evaluation of bias mitigation methods through the creation of AI models with varying degrees of AI bias. The tool consists of two main components:
- Instructions for the implementation of two bias amplification approaches, and
- A graphical user interface (GUI) to assist with the interpretation of the bias mitigation results.
Both bias amplification approaches, (a) quantitative misrepresentation and (b) inductive transfer learning, can be applied to any binary classification model training process, and require only the image data and the associated subgroup and class labels as inputs.
Additional information and example implementations can be found in the User Manual in the GitHub bias.myti.report repository [1].
Intended Purpose
The choice of AI bias mitigation methods can vary between different AI systems. Their effectiveness is dependent on the development and evaluation data sets, model, and classification task. The tool, bias.myti.report has been developed to assist with the evaluation of AI bias mitigation methods’ robustness to bias severity. The tool provides instructions for creating models with varying degrees of bias amplification to evaluate the mitigation method across a wide range of potential bias severities for a specific AI system.
Testing
The bias amplification approaches outlined in bias.myti.report were tested in a simulation study, where the performance bias of a disease status classification model was amplified between subgroups defined by differences in patient demographics [2, 3, 4]. The simulation study demonstrated the amplification approaches’ ability to promote correlation between a specific patient subgroup and disease status. Additionally, the bias.myti.report GitHub repository contains two example Jupyter notebooks demonstrating bias amplication case studies for the different amplification approaches [1].
Limitations
- The tool is only designed to be used for binary image classification AI models.
- Access to the MIDRC data is required to use the example notebooks included with the repository.
- The amplification approaches currently included only amplify AI bias by promoting shortcut learning and do not consider other ways in which bias may present.
- The bias amplification is not fully controlled so experimentation is needed to obtain a specific level of bias in each scenario.
- The tool has only been tested on Linux.
Supporting Documentation
Tool Website:
- Primary: https://github.com/DIDSR/bias.myti.report
- User manual: https://didsr.github.io/bias.myti.report/
References
[1] Y. Zhang, A. Burgon and R. K. Samala, "bias.myti.report," 2024. [Online]. Available: https://github.com/DIDSR/bias.myti.report/tree/main.
[2] A. Burgon, Y. Zhang, N. Petrick, B. Sahiner, K. H. Cha and R. K. Samala, "Bias amplification to facilitate the systematic evaluation of bias mitigation methods," IEEE Journal of Biomedical and Health Informatics, vol. 29(2), 2024. https://doi.org/10.1109/JBHI.2024.3491946
[3] Y. Zhang, A. Burgon, N. Petrick, B. Sahiner, G. Pennello and R. K. Samala, "Evaluation of AI bias mitigation methods by systematically promoting sources of bias," Radiological Society of North America, vol. Physics (Artificial Intelligence in Medical Imaging), no. T6-SSPH08, 2024.
[4] A. Burgon, Y. Zhang, B. Sahiner, N. Petrick, K. H. Cha and R. K. Samala, "Manipulation of sources of bias in AI device development," Proc. SPIE 12927, Medical Imaging 2024: Computer-Aided Diagnosis, 129271J, 2024. https://doi.org/10.1117/12.3008267
Contact
Tool Reference
- RST Reference Number: RST24AI14.01
- Date of Publication: 09/19/2025
- Recommended Citation: U.S. Food and Drug Administration. (2025). bias.myti.report (RST24AI14.01). https://cdrh-rst.fda.gov/biasmytireport