This study conducts the first comprehensive investigation of fairness in 3D medical imaging diagnosis models across multiple protected attributes, including race, gender, and ethnicity. The authors analyze fairness across 2D and 3D models, 5 different architectures, and 3 common eye diseases - age-related macular degeneration (AMD), diabetic retinopathy (DR), and glaucoma.
The results reveal significant biases across demographic subgroups. For example, the White subgroup exhibits improved performance on AMD and DR detection, while the Asian subgroup shows better performance on Glaucoma detection. The Female subgroup exhibits improved performance on AMD detection, while the Male subgroup shows better performance on Glaucoma detection. The non-Hispanic subgroup exhibits improved performance on AMD and Glaucoma detection, while the Hispanic subgroup shows better performance on DR detection.
To address these biases, the authors propose a novel Fair Identity Scaling (FIS) method that incorporates both individual scaling and group scaling to determine loss weights during training. FIS improves both overall performance and fairness, outperforming various state-of-the-art fairness methods.
Additionally, the authors introduce Harvard-FairVision, the first large-scale medical fairness dataset with 30,000 subjects and six demographic identity attributes for eye disease screening, covering three major eye disorders affecting about 380 million people worldwide.
Egy másik nyelvre
a forrásanyagból
arxiv.org
Mélyebb kérdések