Classification images for the detection of a simulated mass in mammographic images

Abstract

 

Mammography is a widely used diagnostic tool for early detection of breast cancer, and visual detection of lesions in a mammogram is a very common visual task in radiology departments. There is significant interest in improving diagnostic performance in mammography through image processing and computer-aided diagnosis. However, it is still not clear at a basic computational level how an observer detects a low contrast lesion that is masked by the presence of normal anatomical structures and quantum noise picked up during image acquisition process.

 

As a first step in examining this question, we use a probit-regression approach to obtain maximum-likelihood estimates of a linear classification image for the task of detecting a low-contrast simulated lesion (5 mm diameter) that has been embedded in mammographic backgrounds derived from a database of digitized mammograms. A total of five subjects participated in 2AFC detection experiments to obtain psychophysical decision data from human observers. We estimate the classification images using probit regression for a set of features that are defined by radial-frequency bands in Discrete Fourier-Transform (DFT) domain. These templates are compared to the templates of a number of linear filter models. Our estimated templates for detecting the mass contain a band of heavily weighted spatial frequencies from 0.08 to 0.3 cycles/mm.