MedPromptX: Grounded Multimodal Prompting for Chest X-ray Diagnosis

Mai A. Shaaban , Adnan Khan , Mohammad Yaqub

Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE

School of Computer Science, Carleton University, Ottawa, CA

---

---

💡 Highlights

• A novel multimodal diagnostic model for chest X-ray images that harnesses multimodal LLMs (MLLMs), few-shot prompting (FP) and visual grounding (VG), enabling more accurate prediction of abnormalities.
• Mitigating of the incompleteness in EHR data by transforming inputs into a textual form, adopting pre-trained MLLMs.
• Extracting the logical patterns discerned from the few-shot data efficiently by implementing a new dynamic proximity selection technique, which allows for the capture of the underlying semantics.

🔥 News

• 2024/03/26: Code is released!
• 2024/05/12: The MedPromptX-VQA dataset is released!

🛠️ Install

Create environment:
conda create -n MedPromptX python=3.8

Install dependencies: (we assume GPU device / cuda available):

cd env

source install.sh

Now, you should be all set.

▶️ Usage

1. Go to scripts/

2. Run:

python main.py --model Med-Flamingo --prompt_type few-shot --modality multimodal --lang_encoder huggyllama/llama-7b --num_shots 6 --data_path prompts_6_shot --dps_type similarity --dps_modality both --vg True

🧳 Checkpoints

Med-Flamingo

OpenFlamingo

LLaMA-7B

✒️ Citation

If you find our work helpful for your research, please consider citing the following BibTeX entry.

@inproceedings{shaaban2025medpromptx,
	title        = {MedPromptX: Grounded Multimodal Prompting for~Chest X-Ray Diagnosis},
	author       = {Shaaban, Mai A. and Khan, Adnan and Yaqub, Mohammad},
	year         = {2025},
	booktitle    = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024 Workshops},
	publisher    = {Springer Nature Switzerland},
	address      = {Cham},
	pages        = {211--222},
	isbn         = {978-3-031-84525-3}
}

♥️ Acknowledgement

Our code utilizes the following codebases: Med-Flamingo and GroundingDINO. We express gratitude to the authors for sharing their code and kindly request that you consider citing these works if you use our code.