# [CVPR 2026 Oral] MEDIC-AD
MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence
CVPR 2026 Oral
---
## ๐ Overview
**MEDIC-AD** is a clinically oriented Vision-Language Model (VLM) designed to bridge the gap between general medical understanding and real-world clinical applications.
It introduces a **stage-wise framework**:
- ๐ **Anomaly Detection** โ lesion-aware representation learning
- ๐ **Difference Reasoning** โ longitudinal symptom tracking
- ๐๏ธ **Visual Explainability** โ clinically grounded heatmaps
This design aligns with real clinical workflows:
**detect โ compare โ explain**
---
## โ๏ธ Installation
We recommend creating a fresh conda environment with Python 3.10.
```bash
conda create -n medic-ad python=3.10 -y
conda activate medic-ad
pip install -r requirements.txt
pip install -e .
```
---
## ๐ Evaluation Datasets
### 1. ๐งช Anomaly Detection (`run_anomaly.py`)
- ๐ฅ Download:
https://drive.google.com/file/d/1TLsgMR6zysq9My57PZF6h8M3bUwgc93n/view?usp=sharing
- Included datasets:
- `Br35H`
- `BrainMRI`
- `HeadCT`
- `ChestX-Det`
- `COVID-19`
#### โถ๏ธ Run (Single GPU)
```bash
python run_anomaly.py \
--model wooohyeooon/MEDIC-AD \
--image-folder /path/to/med_anomaly \
--single_gpu
```
#### โถ๏ธ Run (Multi GPU)
```bash
python run_anomaly.py \
--model wooohyeooon/MEDIC-AD \
--image-folder /path/to/med_anomaly \
--num_gpus 4
```
---
### 2. ๐ฅ Visual Explainability (`run_heatmap.py`)
- ๐ฅ Download:
https://drive.google.com/file/d/1g2x_BUG-Y8zczxCWZEVTAQjO79pNfDbn/view?usp=sharing
- Included datasets:
- `BMAD (BraTS2021, hist-DIY, RESC)`
- `ChestX-Det`
#### โถ๏ธ Run
```bash
python run_heatmap.py \
--model wooohyeooon/MEDIC-AD \
--dataset-roots \
/path/to/med_anomaly_seg/chestx_det/test \
/path/to/med_anomaly_seg/BraTS2021_slice/test \
/path/to/med_anomaly_seg/RESC/test \
/path/to/med_anomaly_seg/hist_DIY/test \
--num_gpus 4
```
---
### 3. ๐ Temporal Reasoning (`run_mmxu.py`)
- ๐ฅ Dataset:
https://huggingface.co/datasets/LinjieMu/MMXU
#### โถ๏ธ Preparation
- Place annotation file as:
```
MMXU-test.jsonl
```
- Set image path to MIMIC-CXR-JPG directory
#### โถ๏ธ Run
```bash
python run_mmxu.py \
--model wooohyeooon/MEDIC-AD \
--image_path /path/to/physionet.org/files/mimic-cxr-jpg/2.1.0/ \
--num_gpus 4
```
---
## ๐ Paper
> **MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence**
> CVPR 2026
๐ Paper: https://arxiv.org/abs/2603.27176 \
๐ Project Page: https://github.com/AIDASLab/Medic-AD
---
## ๐ค Acknowledgement
This repository is built upon:
- [MedEvalKit](https://github.com/alibaba-damo-academy/MedEvalKit)
We also thank the support from:
- NVIDIA AI Technology Center (NVAITC)
- Samsung Changwon Hospital
- Samsung Medical Center
---
## ๐ Citation
If you find this work useful, please cite:
```bibtex
@article{park2026medic,
title={MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence},
author={Park, Woohyeon and Kim, Jaeik and Cho, Sunghwan Steve and Hong, Pa and Jeong, Wookyoung and Nam, Yoojin and Kim, Namjoon and Wong, Ginny Y and Cheung, Ka Chun and Do, Jaeyoung},
journal={arXiv preprint arXiv:2603.27176},
year={2026}
}
```
---