Cerebral Microbleed Detection Model Performance on 758 Patient Scans

758 participant scans were used to train and validate a YOLOv11-based deep learning model for detecting cerebral microbleeds (CMBs) on 2D T2*-weighted GRE MRI. The model achieved a lesion-level F1-score of 0.699 and a patient-level sensitivity of 0.933 for identifying elevated CMB burden. The research, led by Soo-Oh Yang, positions the system as a decision-support tool for ARIA-H risk assessment in Alzheimer's treatment.

Use Cases

• Validate the reported lesion-level sensitivity (0.694) and precision (0.705) metrics for CMB detection performance.
• Analyze the patient-level sensitivity (0.933) and specificity (0.935) for detecting elevated CMB burden (≥4) as a stratification tool.
• Compare model performance on lobar CMBs (0.625 sensitivity) versus deep structure CMBs (0.627 sensitivity) from the regional analysis.
• Reproduce the multi-channel preprocessing pipeline that enhances CMB visibility for 2D T2*-weighted GRE MRI inputs.

Strengths

• Model validation uses a dataset of 758 participants, providing a substantial clinical sample.
• Performance metrics are reported at multiple levels: lesion (F1=0.699), patient (sensitivity=0.933), and region (sensitivity ~0.626).
• Expert consensus served as the reference standard for model training and testing.

Limitations

• The primary data is a 15.9 KB document describing the model, not the underlying image dataset or model weights.
• The dataset size of 758 participants may be limited for training generalizable deep learning models across diverse populations.
• Performance metrics indicate room for improvement, particularly in lesion-level detection (F1-score of 0.699).

Provenance

Source

Soo-Oh Yang via figshare.

Collection Method

Model developed and validated using a dataset of 758 participants, with expert consensus as the reference standard.

Freshness

Last updated March 23, 2026.