The use of contrast for automated pulmonary nodule detection in low-dose computed tomography

A methodology for automated detection of pulmonary nodules in low-dose comp uted tomography (CT) images is investigated. The methodology in question co nsists of (i) identifying some connected set of pixels in the CT image as a potential site for a nodule (this set of pixels is referred to as the targ et); (ii) specifying a further small set of pixels which are believed to be in the healthy tissue surrounding the target (this set of pixels is referr ed to as the background); and (iii) evaluating a function that is calculate d based on the CT values in the target and in the background. The value of this function is indicative of the "noduleness" of the target; hence such a function is referred to as a detector. We present a large family of such d etectors; two widely used detectors included in this family are contrast an d signal-to-noise ratio (SNR). We show that for automated nodule detection contrast appears to be the detector of choice from this family (and, in par ticular, that contrast proves to be a detector superior to SNR for a variet y of situations). Such claims are validated using a receiver operating char acteristic (ROC) study on simulated low-dose CT data sets (representative o f thoracic cross sections with nodules) and with reconstruction algorithms appropriately optimized for the purpose of automated nodule detection. (C) 1999 American Association of Physicists in Medicine. [S0094-2405(99)00603-3 ].