RichardChenZH/MedForge-Reasoner

MedForge-Reasoner: Interpretable Medical Deepfake Detection

MedForge-Reasoner is an 8 billion parameter vision-language model developed by RichardChenZH, based on Qwen3-VL-8B. It is the core model from the MedForge experimental pipeline, specifically engineered for interpretable medical deepfake detection and forgery-aware reasoning.

Key Capabilities

• Medical Image Forgery Detection: Specializes in identifying manipulated regions in medical images, particularly lesion implantation and removal.
• Interpretable Reasoning: Employs a "localize-then-analyze" paradigm, predicting suspicious regions first, then generating an evidence-grounded reasoning chain and final authenticity verdict.
• Visually Grounded Explanations: Provides structured, clinically inspectable reasoning with visual descriptions, localization (bounding boxes), and evidence tied to highlighted regions.
• Broad Modality Coverage: Trained on the MedForge-90K dataset, covering Chest X-Ray, Brain MRI, and Fundus Photography across 19 pathology types.
• Forgery-aware GSPO: Optimized with Forgery-aware Group Sequence Policy Optimization to enhance grounding and reduce visual hallucination.

Intended Use Cases

• Research in medical deepfake detection and multimodal medical forensics.
• Medical image forgery localization and authenticity analysis.
• Benchmarking and ablation studies within the MedForge project.

Limitations

Currently focuses on three 2D imaging modalities, and explanations are in English only. It is intended for research and should not be used for direct clinical diagnosis or autonomous medical decision-making without rigorous validation.