Mgaze-US is a large-scale multi-center, multi-organ ultrasound dataset with gaze annotations.
It is designed to support research on gaze-driven human–robot interaction, visual prompting, medical image segmentation, and robotic ultrasound systems (RUSS).
This dataset has been curated and preprocessed to remove scanner-specific artifacts and metadata, enabling broad usability across domains.
- Total Images: 15,649
- Centers: 8 countries 🌍 across 5 continents
- Modalities: Standard B-mode ultrasound
- Annotations:
- ✅ Manual segmentation labels Y
- 👁️ Gaze fixation positions
- 🔥 Gaze heatmaps
The dataset spans a wide variety of anatomical structures to encourage cross-organ generalization:
- Thyroid — Colombia
- Breast — Egypt, Brazil
- Liver — China
- Kidney — Canada
- Muscle — Netherlands
- Carotid Artery — Poland
- Brachial Plexus — Unknown
- All ultrasound images were preprocessed using EchoLocator to ensure removal of scanner-specific artifacts and metadata.
- Images are aligned and normalized for cross-domain comparison.
Due to size limitations on GitHub, we only provide partial subsets of Mgaze-US, including Kidney, Thyroid, and Liver. For access to the complete dataset, please visit our release on HuggingFace Datasets: 👉 HAN4BME/MGaze-US
For any questions, you can reach me at [[email protected]].
If you use Mgaze-US in your research, please acknowledge this dataset. The associated paper has been submitted to ICRA 2026 and should be cited once it is accepted.