Observer-Driven Texture Analysis in CT Imaging

Abstract

We have created a faster, less computationally intensive method of estimating human observer performance when tasked with detecting a signal in CT images. This is achieved by using a set of images to train the model observer, extract the “most useful” textures in detecting the input signal, and applying those textures to a new set of images and extracting an SNR2-equivalent metric. This is validated by comparing to the Channelized Hotelling Observer (CHO), the field standard for modeling human observer performance, as well as testing on data with known relative performance, which will be elaborated on in the results section. A large merit in this project is that it can be used to help find appropriate radiation dosing per tissue texture to achieve the best differentiation between a signal (such as cancer) and noise or tissue texture, as well as CT image reconstruction algorithm optimization.