MedForge 90K MedForge 90K is a large scale benchmark for interpretable medical deepfake detection under realistic text guided lesion editing on authentic scans. It pairs real clinical style images with high fidelity forged counterparts (lesion implant/edit and removal ), and provides pixel level edit localization plus expert aligned, visually grounded chain of thought supervision for forgery aware reasoning. This resource accompanies the paper MedForge: Interpretable Medical Deepfake Detection via Forgery aware Reasoning (arXiv:2603.18577). What is in the dataset Modalities: brain MRI, chest X ray, and color fundus photography. Forgery types: edit (lesion implantation / pathology editing) and remove (lesion removal), produced by multiple state of the art MMDiT / latent diffusion / instruction following image editors . Supervision signals: Binary authenticity (real vs. deepfake) and forgery type (real / edit / remove). Gold manipulation regions via bounding box masks aligned to edits. Structured reasoning traces in a fixed template (inside ... ), including global image description, optional suspect region with ... , evidence, and conclusion—designed for localize then analyze , clini…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy