Elasticity Training Helps AI Diagnose Breast Cancer

They then used thousands of data inputs derived from the models in order to train a deep convolutional neural network (CNN) to classify tumors as malignant or benign.

A 5-layer CNN was trained with 8,000 samples for heterogeneity, and a 4-layer CNN was trained with 4,000 samples for nonlinear elasticity. When queried on additional synthetic images, the CNNs achieved classification accuracies of 99.7%−99.9%. The researchers then applied the nonlinear elasticity classifier, which was trained entirely using simulated data, in order to classify displacement images obtained from ten patients with breast lesions; the CNN correctly classified eight out of ten cases.

“The general consensus is these types of algorithms have a significant role to play, including from imaging professionals whom it will impact the most,” said senior author Professor Assad Oberai, PhD, of the USC department of aerospace and mechanical engineering. “However, these algorithms will be most useful when they do not serve as black boxes, but instead, a tool that helps guide radiologists to more accurate conclusions.”

Elastography relies on the generation of shear waves determined by the displacement of tissues induced by the force of a focused ultrasound beam or by external pressure. The shear waves are lateral waves, with a motion perpendicular to the direction of the generating force, traveling slowly, and are rapidly attenuated by tissue. The propagation velocity of the shear waves correlates with the elasticity of tissue.

Related Links:
University of Southern California
Rensselaer Polytechnic Institute