Detection of Cerebral Microbleeds in MR Images Using a Single-Stage Triplanar Ensemble Detection Network (TPE-Det)

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuroscience research

Background:

• Cerebral microbleeds (CMBs) are microscopic hemorrhages linked to various neurological conditions.
• Automated CMB detection is challenging due to their size, distribution, and mimicry.
• Existing two-stage methods hinder clinical workflow integration.

Purpose of the Study:

• To develop a clinically feasible, end-to-end CMB detection network using a single-stage, 3D-aware approach.
• Introduce the triplanar ensemble detection network (TPE-Det) by ensembling 2D CNNs across axial, sagittal, and coronal planes.

Main Methods:

• Retrospective analysis of two datasets (DS1: 116 patients, DS2: 58 subjects) with CMBs.
• Utilized 3T field strength and 3D GRE sequence for SWI reconstructions.
• Employed a paired t-test to assess the statistical significance of the ensembling technique (P < 0.05).

Main Results:

• TPE-Det achieved 96.05% sensitivity and 0.88 FPavg on DS1, with statistically significant improvement.
• On DS2 (without retraining), TPE-Det yielded 85.03% sensitivity and 0.55 FPavg.
• The proposed model demonstrated significantly higher precision compared to other methods.

Conclusions:

• Ensembling multidimensional networks significantly enhances CMB detection precision.
• The TPE-Det approach offers a promising advancement for clinical lesion detection.
• This method could improve diagnostic accuracy and patient outcomes in neurological disease management.