A multi-agent system designed to evaluate the performance of medical image classification models, identify model degradation, and, when required, initiate fine-tuning procedures to restore accuracy.

---

🛠️ Installation

ReclAIm is a Python framework built on PyTorch, MONAI, and smolagents. The steps below create an isolated conda environment and install all dependencies with pip. The setup has been tested on Linux with Python 3.12.

1. Clone the repository

git clone https://github.com/eltzanis/ReclAIm.git
cd ReclAIm

2. Create and activate a conda environment

conda create -n reclaim python=3.12 -y
conda activate reclaim

3. Install PyTorch

Install the build that matches your hardware. Both options install the same verified versions (torch==2.12.0, torchvision==0.27.0).

GPU (CUDA) — recommended, default on Linux. The wheel bundles its own CUDA runtime, so only a compatible NVIDIA driver is required (no system-wide CUDA toolkit needed):

pip install torch==2.12.0 torchvision==0.27.0

CPU only — no GPU required:

pip install torch==2.12.0 torchvision==0.27.0 --index-url https://download.pytorch.org/whl/cpu

The framework automatically uses the GPU when torch.cuda.is_available() is True and otherwise falls back to the CPU, so no code changes are needed.

4. Install the remaining dependencies

pip install -r requirements.txt

5. Verify the installation

python -c "import torch; from reclaim_tools import *; print('ReclAIm OK | CUDA available:', torch.cuda.is_available())"

You should see ReclAIm OK | CUDA available: True (or False for a CPU-only install).

Verified dependency versions

Package	Version	Package	Version
torch	2.12.0	scikit-learn	1.8.0
torchvision	0.27.0	imbalanced-learn	0.14.1
monai	1.5.2	pandas	3.0.3
smolagents	1.26.0	matplotlib	3.10.9
litellm	1.87.0	seaborn	0.13.2
numpy	2.4.6	pillow	12.2.0
jupyterlab	4.5.7	ipykernel	7.2.0

---

🚀 Quick start

ReclAIm is driven from the reclaim.ipynb notebook. Launch it inside the activated environment:

jupyter lab reclaim.ipynb

The agent talks to a language model through smolagents' LiteLLMModel, so you need to configure a model and provide the corresponding API key, for example:

from smolagents import LiteLLMModel

model = LiteLLMModel(model_id="openai/gpt-4.1", api_key="YOUR_API_KEY")

Any LiteLLM-supported provider can be used by changing model_id and supplying the matching API key.

---

📚 Cite this work

If you use ReclAIm in your research, please cite:

Eleftherios Tzanis, Michail E. Klontzas (2026). ReclAIm: A Multi-Agent Framework for Monitoring and Correcting Performance Decline in Medical Imaging AI. Radiology: Artificial Intelligence. https://doi.org/10.1148/ryai.250923