Skin Cancer AI Without Source Data

🔥 What happened Researchers dropped CAT-CKD, a method that trains skin cancer models across multiple clinics without ever touching the original source data. Instead of fine-tuning the whole backbone, they train just 4.3M parameters inside a frozen ViT. 💡 Why it matters Privacy rules routinely block data sharing between hospitals — this sidesteps that entirely. It hits 86.1% AUROC across five public datasets, while prior methods either need source data or suffer catastrophic forgetting on tiny clinical sets. ⚡ Our take Finally a paper solving a real regulatory headache instead of chasing leaderboard points. But 86% AUROC is still no dermatologist — clinical deployment remains a stretch.