Radiologists often compare sequential radiographs to identify areas of pathologic change; however, this process is prone to error, as human anatomy can obscure the regions of change, causing the radiologists to overlook pathology. Temporal subtraction (TS) images can provide enhanced visualization of regions of change in sequential radiographs and allow radiologists to better detect areas of change in radiographs. Not all areas of change shown in TS images, however, are actual pathology. The purpose of this study was to create a computer-aided diagnostic (CAD) system that identifies which regions of change are caused by pathology and which are caused by misregistration of the radiographs used to create the TS image. The dataset used in this study contained 120 images with 74 pathologic regions on 54 images outlined by an experienced radiologist. High and low (“light” and “dark”) gray-level candidate regions were extracted from the images using gray-level thresholding. Then, sampling techniques were used to address the class imbalance problem between “true” and “false” candidate regions. Next, the datasets of light candidate regions, dark candidate regions, and the combined set of light and dark candidate regions were used as training and testing data for classifiers by using five-fold cross validation. Of the classifiers tested (support vector machines, discriminant analyses, logistic regression, and k-nearest neighbors), the support vector machine on the combined candidates using synthetic minority oversampling technique (SMOTE) performed best with an area under the receiver operating characteristic curve value of 0.85, a sensitivity of 85%, and a specificity of 84%.
Radiologists frequently use chest radiographs acquired at different times to diagnose a patient by identifying regions of change. Temporal subtraction (TS) images are formed when a computer warps a radiographic image to register and then subtract one image from the other, accentuating regions of change. The purpose of this study was to create a computeraided diagnostic (CAD) system to threshold chest TS images and identify candidate regions of pathologic change. Each thresholding technique created two different candidate regions: light and dark. Light regions have a high gray-level mean, while dark regions have a low gray-level mean; areas with no change appear as medium-gray pixels. Ten different thresholding techniques were examined and compared. By thresholding light and dark candidate regions separately, the number of properly thresholded regions improved. The thresholding of light and dark regions separately produced fewer overall candidate regions that included more regions of actual pathologic change than global thresholding of the image. Overall, the moment-preserving method produced the best results for light regions, while the normal distribution method produced the best results for dark regions. Separation of light and dark candidate regions by thresholding shows potential as the first step in creating a CAD system to detect pathologic change in chest TS images.