Nodule detection in digital chest radiography: effect of anatomical noise

The image background resulting from imaged anatomy can be divided into those components that are meaningful to the observers, in the sense that they are recognised as separate structures, and those that are not. These latter components (reffered to as anatomical noise) can be removed using a method developed within the RADIUS group. The aim of the present study was to investigate whether the removal of the anatomical noise results in images where lung nodules with lower contrast can be detected. A receiver operating characteristic (ROC) study was therefore conducted using two types of images: clinical chest images and chest images in which the anatomical noise had been removed. Simulated designer nodules with a full-width-at-fifth-maximum of 10 mm but with varying contrast were added to the images. The contrast needed to obtain an area under the ROC curve of 0.80, C0.8, was used as a measure of detectability (a low value of C0.8 represents a high detectability). Five regions of the chest X ray were investigated and it was found that in all regions the removal of anatomical noise led to images with lower C0.8 than the original images. On average, C0.8 was 20% higher in the original images, ranging from 7% (the lateral pulmonary regions) to 41% (the upper mediastinal regions).