Generalization of UNET-only bias findings to other segmentation architectures

Determine whether the skin-color bias findings reported in Benčević et al. (2024) for deep learning-based skin lesion segmentation using UNET-based models also hold for other segmentation architectures, such as DeepLabV3 and Vision Transformer-based DINOv2.

Background

The paper reviews prior work on fairness in skin lesion segmentation and notes that the only comprehensive study assessing bias in segmentation models focused exclusively on UNET-based architectures. That study also evaluated skin tone using discrete groupings, which may miss within-image gradations.

Because model architecture can affect feature extraction and error modes, it is important to know whether bias patterns identified in UNET-based models persist in other families such as DeepLabV3 or Vision Transformer-based models like DINOv2. The authors explicitly flag uncertainty about this generalization in the related work section.