Scholarlink

Recent research has revealed the remarkable ability of artificial intelligence (AI) to identify features related to an individual’s self-reported race in medical images, but what such features may be remains an unanswered question. In this work, we aim to identify image regions relevant to race prediction. We argue that previous methods toward this goal (namely, occlusion maps) are not sufficient as they are unable to locate such regions, and we propose to use Cycle-GANs as an alternative. Specifically, we train a Race-specific Cycle-GAN to artificially transform images from patients of one race to images from patients of a different race. We then obtain difference maps by computing the pixel-wise difference between original and transformed image. Difference maps highlight pixels whose values are crucial for an image to be considered as belonging to a patient of a specific race. Additionally, we examine whether such regions are gender dependent by subgrouping our analysis for male and female patients. We show how difference maps are able to identify relevant image regions when previously introduced methods fail, and that, while some differences do exist between genders, the relevant regions mainly overlap.

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

Cycle-GANs Generated Difference Maps to Interpret Race Prediction from Medical Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Lakshika Rathi,
  • Giacomo Nebbia,
  • Ken Chang,
  • Sourav Kumar,
  • Aarushi Gupta,
  • Syed Rakin Ahmed,
  • Jay Patel,
  • Christopher Clark,
  • Yoga Advaith Veturi,
  • Aaron Coyner,
  • Aakanksha Rana,
  • Christopher Bridge,
  • Stephen McNamara,
  • J. Peter Campbell,
  • Matthew Li,
  • Jayashree Kalpathy-Cramer,
  • Praveer Singh

摘要

Recent research has revealed the remarkable ability of artificial intelligence (AI) to identify features related to an individual’s self-reported race in medical images, but what such features may be remains an unanswered question. In this work, we aim to identify image regions relevant to race prediction. We argue that previous methods toward this goal (namely, occlusion maps) are not sufficient as they are unable to locate such regions, and we propose to use Cycle-GANs as an alternative. Specifically, we train a Race-specific Cycle-GAN to artificially transform images from patients of one race to images from patients of a different race. We then obtain difference maps by computing the pixel-wise difference between original and transformed image. Difference maps highlight pixels whose values are crucial for an image to be considered as belonging to a patient of a specific race. Additionally, we examine whether such regions are gender dependent by subgrouping our analysis for male and female patients. We show how difference maps are able to identify relevant image regions when previously introduced methods fail, and that, while some differences do exist between genders, the relevant regions mainly overlap.