HeteroMRI: Robust white matter abnormality classification across multi-scanner MRI data.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Magnetic resonance imaging (MRI) is vital for analyzing brain white matter abnormalities.
• Machine learning integration enhances MRI diagnostics but struggles with heterogeneous, multi-scanner data, especially in rare diseases.
• Developing scanner- and protocol-independent methods is crucial for clinical MRI analysis.

Purpose of the Study:

• Introduce HeteroMRI, a deep learning method for classifying brain MRIs with white matter abnormalities.
• Mitigate the impact of MRI data heterogeneity on classification performance.
• Enhance diagnostic capabilities for white matter abnormalities in diverse clinical settings.

Main Methods:

• Developed HeteroMRI, a deep learning technique utilizing intensity clustering of white matter tissue.
• Applied HeteroMRI to classify MRIs from 11 public datasets encompassing 40 MRI protocols.
• Trained the binary classifier on 200 MRIs to assess performance and generalizability.

Main Results:

• HeteroMRI achieved an average accuracy of 93% ± 4% in classifying white matter abnormalities.
• The method demonstrated robustness in limited data scenarios, simulating rare diseases.
• Accuracy decreased by only 4% and 12% when training data was reduced by 64% and 75%, respectively.

Conclusions:

• HeteroMRI enables classification of heterogeneous MRI data for white matter abnormalities without pre-processing for heterogeneity.
• The method exhibits high independence from MRI scanner and acquisition protocols.
• HeteroMRI shows strong generalizability to unseen MRI protocols, paving new avenues for clinical application.