| # 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](https://arxiv.org/abs/2503.23149), [homepage](https://progemu.github.io/), [model](https://huggingface.co/massaki75/progemu/tree/main)). 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. | |