APPLICATION OF MAMMOGRAPHY PLUS ULTRASOUND BASED RADIOMICS IN THE IDENTIFICATION OF BENIGN AND MALIGNANT BREAST TUMORS

Hezam Mohammed Ahmed Mohammed, Wei Wang*, Hang Qu, Han Yu, Yifan Du and Prosenjit Paul*

ABSTRACT

The aim of the work was to develop and evaluate machine learning models based on characteristics retrieved by conventional radiomics. A total of 92 patients with breast tumors who underwent mammography and ultrasonography were included in the study. Tumors were confirmed by pathological biopsy. The CC and ROI regions were mapped independently by two highly experienced surgeons using ITK-snap software. Data were randomly assigned in a 7:3 ratio to the training cohort and test cohort. LASSO (Least absolute relevance and selection operator) were used to screen, select the features, and develop the radiomics signature. Four machine learning models (RandomForest, DecisionTrees, KNN and Bayes) performance were evaluated with ROC (Receiver Operating Characteristic Curve), and decision curves. The AUC (Area under the ROC curve) of the model, RandomForest, DecisionTrees, KNN and Bayes in the training set and test set were 1.0, 0.969; 0.996, 0.87; 1.0, 0.857; 0.897, 0.895, respectively. AUC values of Tree and Forest model are statistically different at p<0.05. Forest model showed a statistically significant difference with Bayes and Tree model at p<0.05. Our analysis confirms that Forest model outperforms the other three models in terms of AUC score for breast tumor prediction.

Keywords: Breast cancer; machine learning; mammography; features.