Automated Disentangling Analysis of Skin Colour for Lesion Images

A team of researchers has introduced a novel AI framework designed to solve a critical problem in automated dermatology: the performance degradation of machine-learning models when skin color in images (SCCI) differs between training data and real-world deployment. Published on arXiv, the paper 'Automated Disentangling Analysis of Skin Colour for Lesion Images' proposes a method to disentangle the entangled factors of SCCI—such as camera settings, illumination, and intrinsic skin tone—into a structured, manipulable latent space learned from unlabeled images.

The technical core involves two key innovations. First, a 'randomized, mostly monotonic decolourization mapping' prevents information leakage that typically hinders the learning of dark color features. Second, a 'geometry-aligned post-processing step' suppresses unintended color shifts in localized patterns like scars or ink marks during manipulation. This enables precise counterfactual editing, allowing clinicians to visualize what a specific lesion would look like under different skin tones or lighting conditions. It also facilitates controlled traversal along physically meaningful axes like blood perfusion or white balance.

In practice, this framework allows for two major applications: educational visualization of skin conditions across diverse SCCI and robust dataset augmentation for model training. The authors demonstrate that using their method for color normalization and augmentation leads to competitive performance in lesion classification tasks. This addresses a well-known bias in medical AI, where models trained on limited skin-tone datasets often fail to generalize, potentially leading to diagnostic disparities. The work represents a significant step toward more equitable and reliable computer-aided diagnosis in dermatology.