A comparison of supervised machine learning algorithms and feature vectors for MS lesion segmentation using...

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Neuroscience Research

Background:

• Multiple Sclerosis (MS) lesion segmentation in MRI is crucial for understanding disease progression.
• Automated methods are needed to efficiently analyze large medical datasets.
• Current understanding of factors influencing automated lesion segmentation performance is limited.

Purpose of the Study:

• To evaluate the impact of machine learning algorithms and feature extraction on MS lesion segmentation accuracy.
• To identify key drivers of performance in automated lesion segmentation methods.
• To guide the development of more effective MS lesion detection tools.

Main Methods:

• Supervised classification algorithms were trained and validated using manual lesion segmentations.
• Feature extraction functions utilized T1-weighted, T2-weighted, and FLAIR MRI voxel intensities.
• Performance was assessed based on the choice of classification algorithm and feature engineering.

Main Results:

• Differences in feature vectors had a greater impact on segmentation performance than the choice of machine learning algorithm.
• Features incorporating information from neighboring voxels significantly enhanced lesion segmentation accuracy.
• Simple, interpretable algorithms (e.g., logistic regression, LDA, QDA) performed well.

Conclusions:

• Feature engineering is paramount for improving automated MS lesion segmentation.
• Prioritize developing sophisticated features over complex machine learning algorithms.
• Simple algorithms combined with advanced features offer an effective approach for MS lesion segmentation.