Automated lesion detection on MRI scans using combined unsupervised and supervised methods.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning in Medicine

Background:

• Accurate detection of brain lesions on MRI is crucial for correlating lesion location with behavioral impairments.
• Current methods require precise lesion identification for clinical diagnosis and research.

Purpose of the Study:

• To develop a novel, automated method for detecting brain lesions from T1-weighted 3D MRI scans.
• To combine unsupervised and supervised learning approaches for enhanced lesion detection accuracy.

Main Methods:

• Unsupervised methods perform unified segmentation normalization and generate tissue probability maps.
• Atlas-based registration refines probability maps, which are combined with MRI data to create features.
• Supervised methods train Support Vector Machine (SVM) classifiers using these features for voxel-based lesion classification.

Main Results:

• The method achieved a 73.1% Dice coefficient on in-house stroke patient MRIs compared to neurologist delineations.
• On the MICCAI BRATS 2012 dataset, it achieved a 66.5% Dice coefficient for tumor detection.
• Performance was competitive with three state-of-the-art methods on both datasets.

Conclusions:

• A novel automated lesion detection procedure for T1-weighted MRIs was developed.
• The method integrates unsupervised lesion hemisphere identification and probability map refinement with supervised feature extraction and classification.
• This combined approach enables accurate voxel-based lesion classification.