Medtrinity-25m
Massive multimodal medical dataset, containing 25M annotated images on 65 diseases, covering tasks such as report generation or segmentation.
25M images — 18M accessible, image+text formats, multigranular annotations
Not specified (free to access)
Description
Medtrinity-25m is a multimodal medical data set on a very large scale, containing more than 25 million images from 10 different modalities, accompanied by texts, global annotations (type of lesion, modality, inter-regional relationships) and local annotations (bounding boxes, segmentation masks). Three subsets are available, including an accessible subset of 18 million image+text examples, ready to use for training medical AI models.
What is this dataset for?
- Pre-train multimodal foundation models in the medical field
- Develop systems for the automatic generation of radiological reports
- Perform medical vision tasks: classification, detection, segmentation
Can it be enriched or improved?
Yes, the dataset can be enriched by additional annotations, multilingual translations of reports, or an extension to other modalities. It is also possible to filter rare cases or to integrate labels specific to clinical projects. The folder architecture makes it easy to add or modify them in a targeted way.
🔎 In summary
🧠 Recommended for
- Medical AI teams
- Multimodality researchers
- Imaging laboratories
🔧 Compatible tools
- PyTorch
- Hugging Face Datasets
- MONAI
- Segment Anything
- FastMRI tools
💡 Tip
Plan >2TB of storage and plan targeted extractions according to the modality or pathology to avoid memory overloads.
Frequently Asked Questions
What are the concrete use cases with Medtrinity-25m?
It is used to train generative or discriminatory models for medical tasks such as assisted diagnosis, segmentation, or report generation.
Can I use this dataset without medical expertise?
It is not recommended. Medical knowledge is strongly required to interpret annotations and design relevant use cases.
Are there access restrictions or technical limitations?
The accessible subset (18M examples) is freely downloadable, but requires >2TB of disk space and a good hardware configuration.
