Objective  To explore application value of X-ray radiomics in distinguishing benign and malignant breast lesions, and to compare performance of three predictive models.

  Methods  The clinical data of 296 patients with breast lesions was retrospectively analyzed, including 149 cases of malignant lesions and 134 cases of benign lesions. The regions of interest (ROI) of lesions on X-ray images of the craniocaudal (CC) position and the mediolateral oblique (MLO) position were manually separated, radiomics features were extracted and inputted into the machine. Supporting vector machine (SVM), logistic regression (LR), and random forest (RF) were conducted for classification learning. Finally, the Receiver Operating Characteristic (ROC) curve was utilized to calculate the area under curve (AUC) and to compare the performance of different models.

  Results  After feature screening, a total of 9 optimal features were selected, including morphological characteristics(sphericity and major axis length), first-order radiomics features (uniformity and mean absolute deviation), and the high-order radiomics featuresgrey level size zone matrix features(large area low gray level emphasis strength contrast), gray level run length matrix features(intensity) and gray level co-occurrence matrix(contrast, cluster prominence, cluster tendency). The AUC of SVM, LR, RF was 0.820, 0.758, and 0.805, respectively. The difference in AUC between SVM and LR was statistically significant(P < 0.05).

  Conclusion  Radiomics features including first-order radiomics features and high-order radiomics features are extracted by computer analysis. Among them, contrast, cluster prominence, cluster tendency can depict roughness of image texture, better reflect the heterogeneity of breast tumor, and improve the diagnostic accuracy and specificity.