Deep-Learning-Based MRI Microbleeds Detection for Cerebral Small Vessel Disease on Quantitative Susceptibility...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Cerebral microbleeds (CMB) are key indicators of severe cerebral small vessel disease (CSVD).
• Quantitative susceptibility mapping (QSM) and deep learning offer advanced methods for CMB detection in MRI.
• Accurate CMB identification is crucial for understanding and managing CSVD.

Purpose of the Study:

• To develop and validate a two-stage deep learning pipeline for automated CMB detection on QSM images.
• To improve the accuracy and efficiency of CMB identification in patients with CSVD.
• To establish a semi-automated system for assessing CMB location using the Microbleeds Anatomical Rating Scale (MARS).

Main Methods:

• A retrospective study included 393 patients with CSVD, with 1843 CMBs identified.
• A two-stage deep learning pipeline was employed: Stage I used a 2.5D FRST algorithm and a convolutional network; Stage II utilized V-Net for false positive reduction.
• The model was evaluated using sensitivity and positive predictive value (PPV), with external testing on 78 subjects.

Main Results:

• The pipeline achieved high sensitivities: up to 94.9% in Stage I and 93.5% in Stage II.
• Overall sensitivity reached 88.9% with a false positive rate of 2.87 per subject.
• Sensitivities exceeding 85% were reported across nine different brain regions based on the MARS system.

Conclusions:

• A novel deep learning pipeline effectively detects CMB on QSM in a CSVD cohort.
• The proposed method demonstrates superior performance compared to traditional handcrafted approaches for CMB detection.
• The pipeline facilitates a semi-automated MARS scoring system, aiding in the assessment of CMB distribution.