[CVPR 2026] MEDIC-AD

MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence
CVPR 2026

---

📌 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:

• 🔍 Anomaly Detection — lesion-aware representation learning
• 🔄 Difference Reasoning — longitudinal symptom tracking
• 👁️ Visual Explainability — clinically grounded heatmaps

This design aligns with real clinical workflows:
detect → compare → explain

---

⚙️ Installation

We recommend creating a fresh conda environment with Python 3.10.

conda create -n medic-ad python=3.10 -y
conda activate medic-ad

pip install -r requirements.txt
pip install -e .

---

📊 Evaluation Datasets

1. 🧪 Anomaly Detection (run_anomaly.py)

• 📥 Download:
  https://drive.google.com/file/d/1TLsgMR6zysq9My57PZF6h8M3bUwgc93n/view?usp=sharing

• Included datasets:

  • Br35H
  • BrainMRI
  • HeadCT
  • ChestX-Det
  • COVID-19

▶️ Run (Single GPU)

python run_anomaly.py \
  --model wooohyeooon/MEDIC-AD \
  --image-folder /path/to/med_anomaly \
  --single_gpu

▶️ Run (Multi GPU)

python run_anomaly.py \
  --model wooohyeooon/MEDIC-AD \
  --image-folder /path/to/med_anomaly \
  --num_gpus 4

---

2. 🔥 Visual Explainability (run_heatmap.py)

• 📥 Download:
  https://drive.google.com/file/d/1g2x_BUG-Y8zczxCWZEVTAQjO79pNfDbn/view?usp=sharing

• Included datasets:

  • BMAD (BraTS2021, hist-DIY, RESC)
  • ChestX-Det

▶️ Run

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. 🔄 Temporal Reasoning (run_mmxu.py)

• 📥 Dataset:
  https://huggingface.co/datasets/LinjieMu/MMXU

▶️ Preparation

• Place annotation file as:

  MMXU-test.jsonl
  

• Set image path to MIMIC-CXR-JPG directory

▶️ Run

python run_mmxu.py \
  --model wooohyeooon/MEDIC-AD \
  --image_path /path/to/physionet.org/files/mimic-cxr-jpg/2.1.0/ \
  --num_gpus 4

---

📄 Paper

MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence
CVPR 2026

📌 Paper: https://arxiv.org/abs/2603.27176
📌 Project Page: https://github.com/AIDASLab/Medic-AD

---

🤝 Acknowledgement

This repository is built upon:

• MedEvalKit

We also thank the support from:

• NVIDIA AI Technology Center (NVAITC)
• Samsung Changwon Hospital
• Samsung Medical Center

---