# An investigation into the impact of deep learning model choice on sex and race bias in cardiac MR segmentation

**Authors:** Tiarna Lee, Esther Puyol-Antón, Bram Ruijsink, Keana Aitcheson, Miaojing Shi, Andrew P. King  
**arXiv:** [2308.13415](https://arxiv.org/abs/2308.13415)  
**Date:** 2023-08-25  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2308.13415v1/

## Key points

- AI is increasingly used in medical imaging to automate tasks
- Concerns about biases in AI algorithms leading to unequal outcomes among demographic groups
- Study by Tiarna Lee et al. focused on sex and race bias in cardiac MR segmentation
- Imbalances in subject sex and race within training datasets can influence AI-based image segmentation
- Three convolutional neural network-based models and one vision transformer model were evaluated
- Significant sex bias found in three out of four models, racial bias observed across all models
- Severity and nature of biases varied depending on the specific model used
- Importance of selecting appropriate model for fair AI-based segmentation in medical imaging tasks
- Addressing biases through thoughtful model choice can lead to more equitable and accurate medical imaging technologies

## Summary

In the field of medical imaging, artificial intelligence (AI) is increasingly utilized to streamline and automate routine tasks. However, concerns have been raised regarding the potential biases exhibited by these AI algorithms, which can result in unequal performance outcomes among different demographic groups. In a recent study conducted by Tiarna Lee, Esther Puyol-Antón, Bram Ruijsink, Keana Aitcheson, Miaojing Shi, and Andrew P. King, the impact of deep learning model selection on sex and race bias in cardiac magnetic resonance (MR) segmentation was investigated. The researchers specifically focused on how imbalances in subject sex and race within training datasets can influence AI-based cine cardiac MR image segmentation. To assess this phenomenon, three convolutional neural network-based models and one vision transformer model were evaluated. The findings revealed significant sex bias present in three out of the four models examined, while racial bias was observed across all models. Interestingly, the severity and nature of these biases varied depending on the specific model used. This study underscores the critical importance of carefully selecting an appropriate model when training fair AI-based segmentation models for medical imaging tasks. By acknowledging and addressing biases inherent in AI algorithms through thoughtful model choice, researchers can work towards developing more equitable and accurate medical imaging technologies that benefit all individuals regardless of their sex or race.