Probabilistic multiple sclerosis lesion classification based on modeling regional intensity variability and local...

Area of Science:

• Medical imaging analysis
• Neurology
• Machine learning

Background:

• Multiple sclerosis (MS) is a chronic neurological disease characterized by lesions in the brain.
• Accurate classification of MS lesions is crucial for diagnosis and treatment monitoring.
• Existing methods often struggle with lesion variability and challenging brain regions.

Purpose of the Study:

• To present a fully automatic probabilistic method for classifying T1-hypointense and T2-hyperintense MS lesions.
• To improve the accuracy and reliability of MS lesion detection using multimodal magnetic resonance imaging (MRI).

Main Methods:

• Developed a probabilistic framework generating posterior probability density functions for tissues and lesions at each voxel.
• Incorporated spatial variability by segmenting the brain into anatomical regions and building regional likelihood distributions.
• Utilized Markov random fields to ensure local class smoothness by including neighboring voxel information.
• Validated the method on two multisite clinical trial datasets (100 MS patients) comparing results with and without regional information and state-of-the-art techniques.

Main Results:

• The proposed method demonstrated statistically significant improvements in Dice overlap, sensitivity, and positive predictive rates compared to methods without regional information and a widely used approach.
• Enhanced classification accuracy, particularly in challenging areas like the posterior fossa.
• Achieved superior performance in both voxel-based and lesion-based classifications.

Conclusions:

• The developed method provides accurate tissue labels for T1-hypointense and T2-hyperintense MS lesions.
• Offers clinicians a confidence level for classification results, aiding in the assessment of MS.
• Represents a significant advancement in automated MS lesion analysis.