Improving the sensitivity of liver tumor classification in ultrasound images via a power-law shot noise model.

Abstract

Power laws have been observed in various fields and help us understand natural phenomena. Power laws have also been observed in ultrasound images. This study used the power spectrum of the signal identified from the reflected ultrasound signal observed in ultrasonography based on the power-law shot noise (PLSN) model. The power spectrum follows a power law, which has a scaling factor that depends on the characteristics of the tissue in the region where the ultrasound wave propagates. To distinguish between a tumor and blood vessels in the liver, we propose a classification model that includes a scaling factor based on ResNet, a deep learning model for image classification. In a task to classify 6 types of tissue - a tumor, the inferior vena cava, the descending aorta, the Gleason sheath, the hepatic vein, and small blood vessels - tumor sensitivity increased 3.8% and the F-score for a tumor improved 2% while precision was maintained. The scaling factor obtained using the PLSN model was validated for classification of liver tumors.