Over the past decade, radiomic features have proved to be helpful in characterizing tumor biology in vivo by correlating imaging with ground truth pathology. In this study, we identified and utilized such features applied to transmission ultrasound (TU). An abundance of imaging biomarkers are encoded in the TU speed-of-sound maps of breast tissue, which can be used to characterize breast masses. The purpose of this study was to evaluate the efficacy of using these radiomic features to differentiate benign from malignant breast masses.

We randomly selected 90 pathology-proven cases with space-occupying breast masses (49 benign and 41 malignant) from our imaging database. Masses were included in the study if they were able to be segmented using our segmentation algorithm. Radiomic features, including mass irregularity, circularity, and first-order statistics of the pixel distribution, were calculated. T-tests were used to evaluate each feature in its ability to characterize a mass as benign or malignant (p<0.05 considered significant). These features were used in machine learning-based classifiers to differentiate benign from malignant masses.

Both irregularity and circularity proved to be significantly different when comparing benign and malignant masses. Irregularity was measured to be 0.202 ± 0.014 for benign masses and 0.402 ± 0.019 for malignant masses. Similarly, circularity was measured to be 0.788 ± 0.013 and 0.719 ± 0.017, respectively, demonstrating that fundamental morphological features typically used to differentiate benign from malignant masses can also be derived meaningfully from TU imaging. The mode, median and average speed of sound values showed significant differences for benign and malignant mases. Using the two morphological features along with the speed of sound, our algorithm testing found that K-nearest neighbor method with 10-fold cross-validation provided the highest accuracy of 86.7% (ROC-AUC of 0.85).

Our study shows that a range of radiomic features derived from TU can differentiate benign from malignant breast masses. These features may serve as important tools when developing artificial-intelligence-based and computer-aided diagnosis tools for TU.

Radiomics in transmission ultrasound may contribute to decision support to increase precision in the diagnosis and treatment of breast cancer.