Quantitative Assessment of Tissue Stiffness Using Transfer Learning Ultrasound Elastography: A Breast Cancer Phantom Study

In this letter, a phantom study is performed to investigate the feasibility of quantitative tissue stiffness assessment of breast cancer masses using transfer learning ultrasound elastography. Ultrasound elastography is an imaging modality developed to enhance the low specificity of conventional ultrasound, utilizing manual tissue compression, and tracked strain deformation to reconstruct a color-coded qualitative 2-D tissue stiffness map (elastograms). Elastogram assessment in breast cancer screening is based on the working principle that normal tissue is less stiff than benign and malignant breast masses. A transfer learning ultrasound elastography model is developed to classify the breast masses into quantifiable Young's modulus (kilopascals, kP) values. The transfer learning model combines features of images from ultrasound elastography elastograms and from Google's deep learning model AlexNet. The elastogram features used in the training and validation of the transfer learning model were obtained from a CIRS-049 phantom with inclusions having various Young's moduli. The model was evaluated on a CIRS-059 breast phantom with spherical inclusions. Test results show that the transfer learning model yields greater than 88% validation accuracy. Therefore, with the computer-aided tool of transfer learning, ultrasound elastography has the potential of quantitatively differentiating benign and malignant breast masses based on tissue stiffness values.