DFKI project: AI-generated skin cancer
An AI-assisted diagnosis of skin diseases is not yet reliable for all population groups. A DFKI project could provide a remedy.
AI-generated images of skin diseases at the DFKI trade fair stand.
(Image: Robin Ahrens / heise medien)
The German Research Center for Artificial Intelligence (DFKI) presented a project at the Hannover Messe aimed at improving the reliability and fairness of AI-assisted diagnosis of skin diseases. In the “MedGenAI” project, researchers generated images of skin cancer and other skin conditions using AI. The synthetic images can then be used to reduce the bias of diagnostic AI models.
In particular, there is significantly less image material from young people and individuals with darker skin types in the training data used to train the AI models employed in dermatological diagnostics. As a result, the models are biased and recognize skin diseases in these population groups less reliably.
However, especially in medical diagnostics, it is crucial that AI models operate as fairly and reliably as possible. Therefore, efforts are made to reduce the bias explicitly in training datasets, for example, with synthetic data. Synthetic data can be fully or partially AI-generated. Partially synthetic data, for instance, refers to real data enriched with synthetic data.
Latest State of Image Generation
To reduce the bias of AI models in medical diagnostics, researchers from DFKI developed a modern Diffusion Transformer (DiT). This allows for the generation of images of skin diseases. The DiT model used in the “MedGenAI“ project is relatively compact – it can be trained in just one day on a system with eight graphics cards.
Unlike diffusion models used in earlier versions of image generation tools like Stable Diffusion, DALL-E, or Midjourney, diffusion transformers better model spatial dependencies in image compositions. New AI generation tools such as Google's Nano Banana or OpenAI's now discontinued generative video AI Sora therefore increasingly rely on transformer architectures.
From Melanoma to Vascular Injury
The DiT model trained by DFKI scientists can also create so-called counterfactuals. For example, an AI-generated melanoma can be transformed into a vascular lesion, i.e., an injury to blood vessels. These minimal deviations from a disease pattern are hardly recognizable to the naked eye. However, medical AI tools recognize these differences and would consequently provide a different diagnosis.
Using the generated counterfactuals, it is possible to determine how error-prone a medical analysis model is. The insights gained can then be used to improve AI models in dermatological diagnostics.
(rah)
