ProgEmu: Towards Interpretable Counterfactual Generation via Multimodal Autoregression
This repository contains the model weights for the MICCAI'25 paper: Towards Interpretable Counterfactual Generation via Multimodal Autoregression (arxiv, homepage, model). Supported by Shanghai Innovation Institute (SII).
Highlights 💡
Interpretable Counterfactual Generation (ICG): Jointly produces a counterfactual CXR image and a concise interpretation text that pinpoints progression-induced visual changes.
ICG-CXR Dataset: Over 10k longitudinal CXR quadruples (prior image, prompt, subsequent image, interpretation) that supports ICG task.
ProgEmu Framework: A single multimodal autoregressive transformer that generates visual and textual counterfactuals in one forward pass.
