Gastronet-5M Pretrained Models
Welcome to the official repository for the pretrained weights of ResNet50 and VIT-small models pretrained on the Gastronet-5M dataset. These models are intended for researchers and developers working in medical image analysis, specifically in gastroenterological imaging.
GastroNet-5M is the largest publicly available dataset of gastrointestinal endoscopic images to date. It contains 4,820,653 unlabeled images derived from approximately 500,000 unique endoscopic procedures, collected across eight Dutch hospitals between 2012 and 2020. The dataset covers a wide range of endoscopic procedures in the upper and lower gastrointestinal tract, acquired using endoscopy systems from all major manufacturers.
The dataset is designed to accelerate the development of deep learning systems in gastrointestinal endoscopy, with a particular focus on its function as a pretraining dataset. It is expected to contribute to improved diagnostic accuracy, increased robustness to data heterogeneity, and reduced need for scarce annotated data. Images are provided in a standard image format (PNG) and stored and subdivided into zipped folders containing up to 10,000 images each. A represenative subset of 1,000 images is directly availabe for download.
Prior to inclusion, all images were anonymized on-site at each hospital using proprietary software that masked patient-identifying text and metadata. Some images may contain anonymization artifacts. An additional central quality assurance process was performed that included a manual review of all images to remove any remaining irrelevant or potential patient sensitive data.
Access to the dataset can be requested via the following link: https://cortex.thetavision.nl/dataset-provider/listing/1/
Models
This repository contains the following pretrained models:
- ResNet50 on Pretrained with DINOv1 on Gastronet-5M
- ResNet50 on Pretrained with SIMCLRv2 on Gastronet-5M
- ResNet50 on Pretrained with MOCOv2 on Gastronet-5M
- ResNet50 on Initialized with Billion-Scale weights and Pretrained with DINOv1 on Gastronet-5M
- ResNet50 on Pretrained with DINOv1 on Gastronet-1M
- ResNet50 on Pretrained with DINOv1 on Gastronet-200K
- VIT-small on Gastronet-5M
Citation
Researchers using the dataset and the released pretrained weights should cite the following papers respectively:
Jong, M. R., Boers, T. G. W., Fockens, K. N., Jukema, J. B., Kusters, C. H. J., Jaspers, T. J. M., van Eijck van Heslinga, R. A. H., Slooter, F. C., Struyvenberg, M. R., Bisschops, R., van der Putten, J. A., de With, P. H. N., van der Sommen, F., de Groof, A. J., & Bergman, J. J. (2025). GastroNet-5M: A Multicenter Dataset for Developing Foundation Models in Gastrointestinal Endoscopy. Gastroenterology. https://doi.org/10.1053/j.gastro.2025.07.030
Boers, T. G. W., Fockens, K. N., van der Putten, J. A., Jaspers, T. J. M., Kusters, C. H. J., Jukema, J. B., Jong, M. R., Struyvenberg, M. R., de Groof, J., Bergman, J. J., de With, P. H. N., & van der Sommen, F. (2024). Foundation models in gastrointestinal endoscopic AI: Impact of architecture, pre-training approach and data efficiency. Medical Image Analysis, 98, 103298. https://doi.org/10.1016/j.media.2024.103298