Radiomics derived from T2-FLAIR: the value of 2- and 3-classification tasks for different lesions in multiple...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning in Medicine

Background:

• White matter (WM) lesions in relapsing-remitting multiple sclerosis (RRMS) are pathologically diverse, including contrast enhancement lesions (CELs), iron rim lesions (IRLs), and non-iron rim lesions (NIRLs).
• Current radiomics applications using T2-FLAIR imaging for classifying these WM lesions in RRMS are limited, particularly for three-class differentiation.

Purpose of the Study:

• To investigate the efficacy of radiomics models based on T2-FLAIR images for classifying WM lesions into CELs, IRLs, and NIRLs in RRMS.
• To compare the performance of different feature selection and machine learning algorithms for this classification task.

Main Methods:

• A dataset of 875 WM lesions (92 CELs, 367 IRLs, 416 NIRLs) from RRMS patients was analyzed.
• Feature selection was performed using LASSO, reliefF, and mutual information (MI).
• Discrimination models were built using eXtreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM), with performance evaluated by AUC, accuracy, sensitivity, specificity, and precision.

Main Results:

• For two-class classification (IRLs vs. NIRLs), the LASSO classifier with RF model achieved the highest performance (AUC: 0.893).
• For three-class classification (CELs, IRLs, NIRLs), the LASSO with XGBoost model demonstrated superior discrimination (AUC: 0.920, accuracy: 0.796).

Conclusions:

• Radiomics models derived from T2-FLAIR images show significant potential for discriminating between CELs, IRLs, and NIRLs in RRMS.
• These findings suggest that radiomics can aid in the pathological classification of white matter lesions in RRMS.