CMUH SWAN Microbleed Set
China Medical University Hospital SWAN cerebral microbleed dataset
205 clinical GE SWAN brain MRI exams from China Medical University Hospital, Taiwan (104 with microbleeds, 101 without), with voxel-wise microbleed masks by one neurologist. Used for synthetic-to-real microbleed segmentation; shared only through research collaboration.
Overview
This collection of clinical susceptibility-weighted brain MRI comes from China Medical University Hospital in Taichung, Taiwan. Researchers at the hospital's Artificial Intelligence and Robotics Innovation Center gathered it retrospectively under institutional review board approval to test whether cerebral microbleed detectors trained on synthetic lesions transfer to real scans. Microbleed-free exams served as hosts for inserting synthetic microbleeds, and exams with real microbleeds were used for real-data training and testing. The images are not public.
Composition
205 MRI examinations: 104 with at least one microbleed and 101 without. The paper splits them into a training set of 78 positive exams (687 lesions) and 75 negative exams, and a hold-out test set of 26 positive (171 lesions) and 26 negative exams. Positive exams were assigned to the two sets by lesion burden (fewer than 5, 5 to 9, 10 or more microbleeds) so both have a similar mix. The paper does not report patient age, sex, diagnoses or whether any patient contributed more than one exam.
Acquisition
All exams are GE SWAN (susceptibility-weighted angiography) from routine clinical practice on five GE systems: Discovery MR750w, Optima MR450w, Signa HDxt, Signa Architect and Signa Voyager, at 1.5 T and 3 T. In-plane resolution is 0.43 to 0.51 mm (median 0.47 mm) and slice spacing 0.8 to 1.4 mm (median 1.0 mm). The count per field strength is not given.
Annotations
An experienced neurologist drew voxel-wise masks of every microbleed between 2 and 10 mm in diameter on the SWAN images. Negative exams were checked to contain no visible microbleeds and have no masks.
Known limitations
- Not publicly available. The code repository asks interested researchers to contact the corresponding author about collaboration; no data terms are published.
- One annotator, single hospital, one vendor.
- No demographic or clinical information is reported.
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Contrast / sequence
studies, values can overlap
- Susceptibility-weighted 205 100%
Condition
studies, values can overlap
- Cerebral microbleeds 104 51%
Scanner vendor
studies
- GE HealthCare 205 100%
Country
studies
- Taiwan 205 100%
Split
studies
- Training 153 75%
- Test 52 25%
License and access
Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer
Ask the authors or the data holder. No published process or terms, and access is at their discretion
No published license or data use terms
The data holder has published no license and no data use terms. Any use, sharing or commercial right has to be agreed with the data holder, and nothing can be assumed to be allowed. Default copyright and data protection law still apply.
The code repository states that the clinical images are not publicly distributed and points to the corresponding author for research collaborations. No license or data use terms are published; the repository has no license file.
What you can do
- Commercial use Not stated
- Train ML models Not stated
- Create derived data Not stated
- Publish results Not stated
What you can share
- Share the data Not stated
- Share derived data Not stated
- Share trained models Not stated
What you must do
- Cite or credit Not stated
- Share alike Not stated
- Sign an agreement Not stated
- Ethics approval Not stated
- Manuscript review Not stated
- Release code Not stated
- Return results Not stated
- Delete after use Not stated
Limits
- No re-identification Not stated
- Location limits Not stated
Citation
Hsu T-L, Lai T-Y, Lin C-T, Huang C-H, Wang W-C. Synthetic-to-Real Transfer in Cerebral Microbleed Generation and Segmentation. arXiv:2610.00743 (2026).
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Studies | total MRI examinations; the number of distinct patients is not reported | 205 | hsu2026 Section 2.1 |
| Studies | contrast=swi GE SWAN | 205 | hsu2026 Section 2.1 |
| Studies | condition=cerebral_microbleeds CMB-positive; 101 exams are CMB-negative | 104 | hsu2026 Section 2.1 |
| Studies | vendor=ge Discovery MR750w, Optima MR450w, Signa HDxt, Signa Architect, Signa Voyager | 205 | hsu2026 Section 2.1 |
| Studies | country=TW China Medical University Hospital | 205 | hsu2026 Section 2.1 |
| Studies | split=train 78 CMB-positive (687 lesions) and 75 CMB-negative; summed from Table 1, which is disjoint and complete | 153 | hsu2026 Table 1 |
| Studies | split=test hold-out test set: 26 CMB-positive (171 lesions) and 26 CMB-negative | 52 | hsu2026 Table 1 |
| Studies | split=train;condition=cerebral_microbleeds | 78 | hsu2026 Table 1 |
| Studies | split=test;condition=cerebral_microbleeds | 26 | hsu2026 Table 1 |
Sources
The keys used in the table above.
- hsu2026 Hsu et al. 2026, arXiv:2610.00743v1 paper
- hsu2026-repo synthetic-to-real-cmb-detection GitHub repository, README (Data Availability) website