Selection of Radiomics Features based on their Reproducibility

Area of Science:

• Radiomics
• Machine Learning
• Clinical Data Analysis

Background:

• Machine learning in clinical settings faces challenges with small, unbalanced datasets, leading to artifacts.
• Current dimensionality reduction methods often overlook feature repeatability and uncertainty.
• Radiomics features can be sensitive to variations in image acquisition and physician annotations.

Purpose of the Study:

• To propose a novel feature selection strategy for radiomics using reproducibility.
• To enhance the performance of machine learning models in clinical applications with small sample size (SSS) datasets.
• To identify radiomics features with high inter-class correlation coefficient (ICC) across different variability sources.

Main Methods:

• Investigated reproducibility of radiomics features across three studies focusing on convolution kernel, image acquisition parameters, and inter-observer variability.
• Selected features with an ICC greater than 0.7 in all three reproducibility studies.
• Evaluated selected features for lesion malignancy classification on an independent dataset.

Main Results:

• Features selected based on reproducibility demonstrated superior performance in lesion malignancy classification.
• The proposed method outperformed established techniques like Principal Component Analysis (PCA), Kernel Discriminant Analysis via QR decomposition (KDAQR), LASSO, and a custom Convolutional Neural Network.
• Reproducibility-based feature selection effectively addressed limitations of existing dimensionality reduction approaches.

Conclusions:

• Radiomics feature reproducibility is a critical factor for robust machine learning in clinical SSS unbalanced datasets.
• The proposed ICC-based selection method offers a significant improvement over current state-of-the-art techniques.
• This approach enhances the reliability and accuracy of radiomics in medical image analysis and clinical decision-making.