AI Tools Developed at the ATI Lab

At the ATI Lab, we develop robust, clinically meaningful AI tools that bridge cutting-edge methodology with real-world medical imaging and multi-omics applications. Our work focuses on transforming advanced AI research into transparent, reproducible, and deployable solutions for clinical care, biomedical research, and health systems.

Our tools are designed with clinical relevance, regulatory awareness, and translational impact at their core, supporting applications across radiology, oncology, musculoskeletal imaging, and multimodal precision medicine.

mAIstro

mAIstro is an autonomous, open-source multi-agent system designed to orchestrate the full pipeline of medical imaging AI development - from exploratory data analysis (EDA) and radiomics feature extraction to training and deploying deep learning models.

Built around a team of specialized agents, mAIstro enables researchers and clinicians to interact with complex AI workflows using natural language prompts - no coding required.

🌐 LLM-Agnostic Design: mAIstro can operate with both open-source and commercial LLMs (e.g., GPT-4, Claude, DeepSeek, LLaMA, Qwen), providing flexibility across environments.

✨ Key Features

🔎 Autonomous Exploratory Data Analysis (EDA)
🧬 Radiomics feature extraction (for CT, MRI, and multi-parametric imaging)
⚙️ nnU-Net Agent for segmentation model development and implementation
⚙️ TotalSegmentator Agent for full-body and organ-specific automatic segmentation
🩻 Image Classification Agent (ResNet, VGG16, InceptionV3 architectures)
📊 Feature Importance and Feature Selection
📈 Tabular data Classification and Regression Agents
🛠️ Modular tool-based architecture for extensibility
🧾 Integrated in a single user-friendly Jupyter Notebook

mAIstro has been developed by our post-doctoral researcher Eleftherios Tzanis. The journal publication of mAIstro can be found here. To download mAIstro as a standalone desktop app or to use in google colab click on the image below to be redirected to the mAIstro github page.

IW to T2*GRE converter

This Generative Adversarial Network is a first-of-its-kind machine learning model that synthesizes T2* W Gradient Echo from Intermediate Weighted fat-suppressed images of the knee. It is based on the cycleGAN architecture and has been trained on a dataset consisiting of knee examinations from 820 patients (15000+ images).

🔧 What this model does

🧠 Generates synthetic T2*-weighted (GRE/MEDIC) knee MRI images from routinely acquired fat-suppressed intermediate-weighted (FS-IW) sequences
⚡ Eliminates the need for additional T2* acquisitions, avoiding patient recall, extra scan time, and workflow disruption
🩻 Preserves diagnostic information relevant for cartilage assessment, synovial blooming (e.g. PVNS, hemosiderin), chondrocalcinosis, meniscal pathology, and joint effusion
⏱️ Produces images in ~1 second, compared to ~10 minutes of conventional acquisition
🖥️ Works as standalone software via an easy-to-use graphical interface or command-line pipeline
🔓 Fully open-source, with code, trained model, and executable available for clinical and research use

✨ What’s novel

🚀 First open-source deep learning model specifically designed to synthesize T2*-weighted knee MRI sequences
🧬 Trained on large, real-world clinical data (>12,000 images), not curated or synthetic datasets
🔁 CycleGAN-based bidirectional translation, ensuring structural consistency and realistic image appearance
✅ Demonstrated diagnostic interchangeability with conventional T2* images (ICC > 0.95 across findings)
📊 Extensive multi-level validation, including image quality metrics, voxel-level signal analysis, reader studies, and artifact impact assessment
⚠️ Transparent artifact analysis, explicitly identifying when and how synthetic artifacts may occur—and showing minimal diagnostic impact in most cases
🌍 Designed for real-world adoption, lowering barriers for validation, reuse, and further development by the imaging community

This model has been developed by our PhD student Konstantinos Vrettos. The journal publication can be found here and the code, standalone and web-based app can be found by clicking on the image below

Page updated

Google Sites

Report abuse