Comparison of radiomic features of benign and malign breast masses on contrast enhanced mammography images

Purpose: The aim of this study is to extract radiomic features from tissue analysis in low-energy and recombined images of contrast-enhanced mammography (CEM), which can be used for the benign-malignant differentiation of breast masses, and to demonstrate the diagnostic effectiveness of radiomic analysis. Material and Method: After obtaining ethical committee approval, a total of 145 cases were included in the study, between September 2013 and August 2022, who underwent contrast-enhanced mammography (CEM) for diagnostic purposes due to suspicious clinical or radiological findings, and subsequently received a definitive diagnosis. The contrast-enhanced mammography images were acquired using the GE Senographe Essential full-field digital mammography system and evaluated by two radiologists with 20 and 3 years of experience in breast imaging, respectively, at the mammography workstation. Patient images were loaded into the open-source image processing tool ITK-SNAP in DICOM format. For radiomic analysis, the boundary of each lesion was manually delineated to include only the mass. This segmentation process was applied to both low-energy and recombined images. Subsequently, 102 radiomic features were extracted from the two-dimensional tumors obtained from these mammography images using the PyRadiomics Python program. The radiomic features obtained in the study and the matrices used are shape-based features, first-order features, Gray Level Co-Occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Dependence Matrix (GLDM), and Neighboring Gray Tone Difference Matrix (NGTDM). To reduce the dimensionality of the features, the MRMR, ReliefF, and ANOVA algorithms available in MATLAB software were used. Each algorithm generated an importance score and ranking for each feature, and the top 10 features with the highest scores were selected for each algorithm. The dataset was randomly divided into 75% for training and 25% for testing. Different classifiers were trained to distinguish between benign and malignant ROIs by developing a sequential forward feature selection algorithm that selects feature subsets in different ways. For supervised machine learning, ensemble learning, decision trees, naive Bayes, support vector machines, and neural networks algorithms were used. The classification process was performed using 10-fold cross-validation. In model optimization, the Grid search optimization method was used to find the best hyperparameter values by trying all hyperparameter combinations. For test performance evaluation, the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated from the confusion matrix separately for low-energy and recombined images. The best model was selected based on the highest accuracy and AUC value. MATLAB R2022b software (MathWorks, Inc., Natick, MA, USA) was used for all analyses. Results: Within the scope of the study, 44.5% (n=73) of all evaluated masses were benign, and 55.5% (n=91) were diagnosed as malignant. In the benign group, the mean age of patients was 48.1±9.5 years, while in the malignant group, it was 49.9±10.5 years. The mean tumor size was 23.2±20.6 mm in the benign group and 35.7±21.5 mm in the malignant group. Based on the analysis of three different algorithms, the most valuable 22 features were selected for recombined images and the most valuable 25 features for low-energy images. The best classifier for both types of images was ensemble learning. The highest accuracy and AUC values were achieved with ensemble learning for recombined images, which were 91.8% and 0.978, respectively, and for low-energy images, they were 89.7% and 0.968, respectively. The classifiers with the highest sensitivity for recombined images were ensemble learning and support vector machines, both achieving 91.8%. The highest specificity, on the other hand, was obtained with neural networks classifier, which was 95.8%. For low-energy images, the classifier with the highest sensitivity was ensemble learning with 98.0%, while the highest specificity was achieved with support vector machines at 91.7%. Discussion: The results of our study demonstrate that there are discriminative radiomic features that can assist in the benign-malignant differentiation of masses in contrast-enhanced mammography. Furthermore, our findings are promising and significant in terms of providing data for the current topic of artificial intelligence-based radiological automated decision support systems.