Detecting outlying subjects in high-dimensional neuroimaging datasets with regularized minimum covariance determinant

Area of Science:

• Medical Imaging Analysis
• Statistical Data Science
• Biomedical Research

Background:

• Medical imaging datasets are often highly variable and complex.
• Identifying and excluding outlier data is crucial for reliable clinical studies and research.
• Traditional statistical methods struggle with high-dimensional, noisy medical imaging data.

Purpose of the Study:

• To develop a statistically sound method for outlier detection in medical imaging.
• To address the limitations of informal and traditional statistical approaches in high-dimensional settings.
• To improve the quality and reliability of medical imaging datasets used in research.

Main Methods:

• Modification of the classical Minimum Covariance Determinant (MCD) approach.
• Inclusion of a regularization term to ensure well-posed estimation.
• Application to simulated and real-world medical imaging data, including functional brain images.

Main Results:

• The modified MCD approach successfully detects outliers in high-dimensional data.
• The method performs satisfactorily even when the number of dimensions exceeds the number of observations.
• Demonstrated effectiveness on both simulated and real medical imaging datasets.

Conclusions:

• The proposed regularized Minimum Covariance Determinant method offers a robust solution for outlier detection in medical imaging.
• This technique enhances the statistical validity of group studies by ensuring data quality.
• The method is particularly valuable for high-dimensional and noisy functional brain imaging data.