Objectives
To develop radiomic models of B‐mode ultrasound (US) signatures for determining the origin of primary tumors in metastatic liver disease.
Methods
A total of 254 patients with a diagnosis of metastatic liver disease were included in this retrospective study. The patients were divided into 3 groups depending on the origin of the primary tumor: group 1 (digestive tract versus non–digestive tract tumors), group 2 (breast cancer versus non–breast cancer), and group 3 (lung cancer versus other malignancies). The patients in each group were allocated to a training or testing set (a ratio of 8:2). The region of interest of liver metastasis was determined through manual differentiation of the tumors, and radiomic signatures were acquired from B‐mode US images. Optimal features were selected to develop 3 radiomic models using multiple‐dimensionality reduction and classifier screening. The area under the curve (AUC) of the receiver operating characteristic curve was applied to assess each model's performance.
Results
A total of 5936 features were extracted, and 40, 6, and 14 optimal features were sequentially identified for the development of radiomic models for groups 1, 2, and 3, respectively, with training set AUC values of 0.938, 0.974, and 0.768 and testing set AUC values of 0.767, 0.768, and 0.750. The differences in age, sex, and number of liver metastatic lesions varied greatly between the 4 primary tumors (P < .050).
Conclusions
B‐mode US radiomic models could be effective supplemental means to identify the origin of hepatic metastatic lesions (ie, unknown primary sites).