Scholarlink

Deep learning (DL) algorithms have shown great promise in the medical field but face criticism for reflecting and perpetuating systemic biases in healthcare. This study addresses algorithmic biases in ophthalmology, focusing on the effects of demographic attributes like age and gender in DL analyses for eye disease classification. Utilizing the Ocular Disease Recognition (ODIR) dataset of ocular fundus images and demographic data, we explore two approaches for mitigating algorithmic bias: fairness through unawareness and fairness through awareness. We propose using the Domain-Adversarial Neural Network (DANN) approach to create agnostic feature representations, reducing the impact of sensitive demographic attributes on model predictions. The project is conducted in two phases: first, by comparing models trained with and without demographic data to identify biases and their effects; and second, by implementing the DANN-based model to mitigate bias. Our findings demonstrate that bias-aware strategies, combined with robust architectural solutions, can lead to fairer and more reliable AI tools in healthcare. This research highlights the importance of ensuring equitable and high-quality AI-based healthcare solutions for all patient groups.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Domain-Adversarial Neural Networks to Explore Biases in the Diagnosis of Multiple Eye Conditions from Fundus Image Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Chiara Pullega,
  • Arianna Dagliati,
  • Allan Tucker

摘要

Deep learning (DL) algorithms have shown great promise in the medical field but face criticism for reflecting and perpetuating systemic biases in healthcare. This study addresses algorithmic biases in ophthalmology, focusing on the effects of demographic attributes like age and gender in DL analyses for eye disease classification. Utilizing the Ocular Disease Recognition (ODIR) dataset of ocular fundus images and demographic data, we explore two approaches for mitigating algorithmic bias: fairness through unawareness and fairness through awareness. We propose using the Domain-Adversarial Neural Network (DANN) approach to create agnostic feature representations, reducing the impact of sensitive demographic attributes on model predictions. The project is conducted in two phases: first, by comparing models trained with and without demographic data to identify biases and their effects; and second, by implementing the DANN-based model to mitigate bias. Our findings demonstrate that bias-aware strategies, combined with robust architectural solutions, can lead to fairer and more reliable AI tools in healthcare. This research highlights the importance of ensuring equitable and high-quality AI-based healthcare solutions for all patient groups.