A Novel Data Augmentation Method for Radiomics Analysis Using Image Perturbations.

Area of Science:

• Medical imaging analysis
• Radiomics and quantitative imaging
• Machine learning in healthcare

Background:

• Radiomics analysis often faces challenges with imbalanced or small datasets.
• Current methods like over-sampling are applied directly to extracted features, potentially limiting their effectiveness.
• Cardiac imaging analysis requires robust methods for distinguishing between similar conditions like cardiac amyloidosis, aortic stenosis, and hypertrophic cardiomyopathy.

Purpose of the Study:

• To propose and evaluate a novel data balancing and augmentation technique for radiomics using region of interest (ROI) perturbations.
• To apply this perturbation-based method to cardiac computed tomography images for improved classification of cardiac diseases.
• To assess the effectiveness of ROI perturbations in addressing data limitations in radiomics for clinical applications.

Main Methods:

• A novel technique involving perturbations (erosion, dilation, contour randomization) applied to ROIs in cardiac CT images was developed.
• Radiomic features were extracted from both original and perturbed ROIs to create augmented datasets.
• The perturbation-based method was compared against random over-sampling, ADASYN, and SMOTE for balancing and augmenting datasets in classification tasks using support vector machines.

Main Results:

• The perturbation-based approach demonstrated superior performance in classifying cardiac amyloidosis versus aortic stenosis (f1 score: 80%, AUC: 0.91) and hypertrophic cardiomyopathy (f1 score: 86%, AUC: 0.92).
• Feature selection methods including LASSO and PCA were employed to analyze robustness, redundancy, and relevance.
• The technique effectively addressed data balancing and augmentation challenges, leading to improved classification metrics.

Conclusions:

• Region of interest (ROI) perturbations offer a powerful and effective strategy for data balancing and augmentation in radiomics.
• This novel approach enhances the quantitative characterization of medical images, particularly in challenging cardiac disease classifications.
• The findings suggest significant potential for ROI perturbation techniques in improving the reliability and accuracy of radiomics-based diagnostic tools.