Wszystko, co robimy,robimy otwarcie.
Narzędzia, zbiory danych i modele tworzone razem z inżynierami, klinicystami i badaczami. Korzystaj, zgłaszaj issues, wysyłaj PR-y — przyjmujemy nowe osoby na każdym poziomie.
UMIE Datasets
Open-source pipelines that standardize 880k+ images across 20+ medical imaging datasets (CT, MRI, X-ray) into a unified format with RadLex-compliant labels.
Zobacz projekt →
RocketRAG
A speed-focused Retrieval-Augmented Generation framework packaging document ingestion, semantic chunking, vector storage, and LLM inference into a pluggable CLI and FastAPI toolkit.
Zobacz projekt →
LOMA — Offline Medical AI Assistant
A zero-cloud mobile medical assistant that runs the full pipeline — embeddings, retrieval, and language model responses — entirely on the user's phone.
Zobacz projekt →
ColorNephroNet
A two-stage CNN pipeline that colourises grayscale kidney CT scans into pseudo-RGB images to improve malignancy prediction by leveraging features learned from colour image datasets.
Zobacz projekt →
Polish Radiology Report IE
An automatic parametrization model for Polish radiology reports using deep language models, trained on 1,200 annotated chest CT reports labeled with 44 observation tags and achieving an F1 score of 81%.
Zobacz projekt →
SARS-CoV-2 from Blood Count
Machine learning models that detect SARS-CoV-2 infection from routine complete blood count (CBC) tests, offering a cost-effective, scalable screening tool to complement RT-PCR.
Zobacz projekt →
Catheter Injury Prediction
Machine learning models that predict catheter-induced coronary and aortic dissections from clinical, anatomical, and procedural data across more than 80,000 catheterized patients.
Zobacz projekt →
RAD-SRAC
A training-free, image-based retrieval-augmented classification framework that enhances any vision-language model for radiology with few-shot examples and minimal resources.
Zobacz projekt →
Cross-Adaptation
A target-free domain adaptation method that lets machine learning models generalize across datasets from different sources without requiring labeled examples from the target domain.
Zobacz projekt →