SHIVA-CMB: a deep-learning-based robust cerebral microbleed segmentation tool trained on multi-source T2*GRE- and...

Area of Science:

• Neuroimaging
• Radiology
• Artificial Intelligence

Background:

• Cerebral microbleeds (CMB) are key indicators of cerebral small vessel disease (cSVD), linked to cognitive decline, dementia, and stroke.
• Automated detection of CMB using deep learning (DL) is advancing, but a lack of shared pre-trained models limits their practical use and validation.

Purpose of the Study:

• To introduce the SHIVA-CMB detector, a 3D Unet-based tool for automated CMB detection.
• To address the need for accessible and validated DL models for CMB analysis in research settings.

Main Methods:

• Developed a 3D Unet-based deep learning tool (SHIVA-CMB).
• Trained the model on 450 MRI scans from six diverse cohort studies (T2*- and susceptibility-weighted imaging).
• Evaluated performance on held-out test sets and an independent multi-center cohort, comparing against expert ratings.

Main Results:

• Achieved high sensitivity (0.67), precision (0.82), and F1 score (0.74) on a held-out test set with low false positives.
• Demonstrated robust performance on unseen data and strong correlation (R=0.89) with expert visual CMB counts in a large cohort.
• Publicly released the SHIVA-CMB pipeline and pre-trained models to the research community.

Conclusions:

• The SHIVA-CMB detector is a generalizable and accurate tool for automated CMB detection.
• Public sharing of the tool and models facilitates broader research into cSVD pathophysiology and consequences.
• Enables rapid characterization of CMB in large-scale studies, accelerating neuroimaging research.