Med-DLE: Agentic Medical Deep Learning Engineer

MedDLE is designed to be a codegen agentic system for medical deep learning tasks.

Quick Start

Setup environment

1. we recommand use conda to create a virtual environment:

conda create --name meddle python=3.11 
conda activate meddle

2. Then install meddle via:

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

3. remember to set up your private api key of OPENAI/ANTROPIC/OPENROUNTER for openai-compatible endpoint

export OPENAI_BASE_URL="<your base url>" # (e.g. https://api.openai.com/v1)
export OPENAI_API_KEY="<your key>"

Setup knowledge base for RAG

You can create the knowledge base with create_monai_knowledge_base.sh or download from link.

You could use this command to check whether the creation succeeds.

python -m meddle.monai_rag.query_rag_db

Setup dataset

1. Get data from link and move the data into med_dl_tasks. This zip file contains both MedMNIST only.
2. The goals and evaluation targets for each dataset can be found in the corresponding shell scripts:
   • For MedMNIST datasets, see run_medmnist_all.sh.
   • For MCO (Molecular and Cellular Oncology) datasets, see run_mco_all.sh.

Try MedDLE

meddle data_dir=<your dataset> \
     goal=<your goal> \
     eval=<your metrics> \
     agent.steps=30 agent.enable_monai_knowledge_base=true

here's a detailed example:

meddle data_dir="meddle/example_tasks/retinamnist_224" exp_name="torch-example_task_retina"\
     goal="Predict the category of each given Fundus Camera image by selecting the most appropriate one from the 5 available classes." \
     eval="Use the accuracy and the area under curve (AUC) between the predicted class and ground-truth class on the test set." \
     agent.step_plan.proposal=20 agent.step_plan.tuning=10 agent.step_plan.hpo=5

You can set general steps (MedDLE with auto-schedule steps for multiple agents) or detailed step_plan for each agent.

Advanced Features

RAG Mode

We support both basic RAG and advanced (query2doc) mode. They are controlled with the following configs.

agent.enable_monai_knowledge_base=true agent.enable_query2doc=true

Time Limits

MedDLE supports configurable timeouts for code execution. The default timeout is 3600 seconds, which can be adjusted via:

• exec.timeout: Base timeout for all agents (default: 3600s)
• HPO agent automatically gets 2x timeout: 7200s when running hyperparameter optimization

To set a total time limit for all steps, use agent.total_time_limit:

meddle data_dir="meddle/example_tasks/retinamnist_224" exp_name="test_timelimit-example_task_retina"\
     goal="Predict the category of each given Fundus Camera image by selecting the most appropriate one from the 5 available classes." \
     eval="Use the accuracy and the area under curve (AUC) between the predicted class and ground-truth class on the test set." \
     agent.force_monai_with_prompt=false agent.enable_monai_knowledge_base=false agent.expose_prediction=true \
     agent.steps=100 agent.total_time_limit=60

🙏 Acknowledgement

• We thank all medical workers and dataset owners for making public datasets available to the community.
• Thanks to the open-source of the following projects, our code is developed based on their contributions:
  • aideml
  • monai
  • AutoGluon