Addressing fairness and data limitations in dermatological diagnosis through color-invariant representation learning and synthetic data generation

While deep learning-based approaches have demonstrated expert-level performance in dermatological diagnosis tasks, they rely on a data-driven learning paradigm that requires large-scale annotated data and mimic the biases therein (e.g., biases towards skin types). Furthermore, existing public dermatological datasets have limitations such as small size, narrow disease coverage, insufficient annotations, and non-standardized image acquisitions. In this thesis, we propose CIRCLe, a skin color-invariant deep representation learning method for improving fairness in skin lesion classification by utilizing a regularization loss to encourage images with the same diagnosis but different skin types to have similar latent representations. Moreover, we introduce DermSynth3D, a novel framework for synthesizing large-scale densely annotated in-the-wild dermatological images by blending skin disease patterns onto 3D textured meshes of human subjects using a differentiable renderer and generating 2D images from various camera viewpoints under chosen lighting conditions in diverse background scenes.