Application of radiomics features selection and classification algorithms for medical imaging decision: MRI radiomics breast cancer cases study

Abstract

The accurate imaging classification of breast lesion is a capital step for diagnosis and appropriate treatment. The radiomics approach aims to extract and analyze quantitative multiple informative data from medical images unseen by the expert eyes, thus the need for enhancing the ability of the machine learning algorithm to classify different lesions. Generally the radiomics techniques have led to high dimensionality, which may cause overfitting, increasing the complexity of the predictive model and degrading the prediction performance. Thereby, feature selection aims at solving this problem by selecting the relevant features and removing the irrelevant and redundant ones. In this study we present the exploitation of heterogeneous radiomics features for supporting decision-based clinical tasks and the ability of various machine learning classifiers, combined with several feature selection methods, to accurately predict breast cancer nodules. This study uses 593 radiomics features to quantify the intensity, shape and texture of breast tumor images to classify the binary outcome of the benign or malignant status. The uncertainty feature selection, in addition to the decision tree have scores the best performance among 66 other feature selection/classifier combinations (accuracy = 0.85). On the other hand, the t-Test and the Gaussian naïve Bayes have the worst performing combination (accuracy = 0.49). This study proves also that morphological features influence the behavior of machine learning algorithms.

Keywords

• Feature selection
• Machine learning algorithm
• Classification
• Medical decision making
• Radiomics
• Breast cancer