Synthetic microbleeds generation for classifier training without ground truth.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Cerebral microbleeds (CMB) are key indicators of cerebrovascular diseases and cognitive decline.
• Susceptibility Weighted Imaging (SWI) on MRI detects CMB as hypointense lesions.
• Manual CMB detection is time-consuming due to low prevalence and mimics, hindering automated methods.

Purpose of the Study:

• To develop a synthetic CMB (sCMB) generation model for training and testing automated detection systems.
• To create a large, diverse dataset of realistic synthetic ground truth for machine learning.
• To improve the accuracy and efficiency of CMB classification using supervised methods.

Main Methods:

• Modeled sCMB using random Gaussian shapes and integrated them into healthy brain MRI data.
• Compared training machine learning models on synthetic data versus standard augmentation techniques.
• Validated the sCMB model through whole-brain detection experiments using a 10-fold cross-validation with an ensemble of neural networks.

Main Results:

• Training a random forest model solely on synthetic data achieved near state-of-the-art performance (~9 false positives per scan) when tested on real lesions.
• Demonstrated that optimal detection performance can be achieved using synthetic CMB for training.
• A dataset with 37,000 synthetic lesions is publicly available for benchmarking and training.

Conclusions:

• The proposed synthetic microbleed model serves as a powerful data augmentation strategy for CMB classification.
• This approach should be considered for developing robust automated lesion detection systems in SWI MRI.