Mammographic computer-aided detection using bootstrap ensembles

Breast cancer is a leading cause of morbidity and mortality in the U.S. Screening mammography has contributed to their amelioration through earlier detection, increasing the chances for a more effective treatment. Computer-Aided Detection has proven to be a useful supplement to the radiologist's assessment.; A method for the statistical segmentation of potential breast lesions is presented in this context of computerized detection. It is based on the construction of ensembles of circular templates fitted to bootstrap samples of mammographic images. Three bootstrapping schema are proposed for resampling data with an inherent spatial correlation structure, residuals, a moving window, and standardized window. Also, three different model combination methods are proposed, conventional bagging, false detection rate bagging, and “at least one” vote. The methods are tested on sets of simulated and real images to examine the statistical and practical performance of the methods. This performance is assessed in terms of misclassification rate, sensitivity, and specificity, on a pixel basis obtained from class probability contours. Additionally, a bias-variance decomposition allows to determine the roots of improvement of the ensemble classifier over the individual classifier. Comparisons are made with two other methods, the Average Fraction Under the Minimum (AFUM) filter and a Markov Random Field (MRF) model.; The major findings are: (1) the circular template performs better than the average segmentation using MRF. However, it may present difficulties when dealing with patient images that do not conform well to the template. (2) The combination methods establish a continuum in the operative point of the classifier operation, with majority vote in the side of high specificity and “at least one” in the end of very high sensitivity. (3) Bagging and MRF were found to indicate comparable class labels. (4) Histogram equalization improved the detection rate for patient images. (5) Bagging approaches decreased variance, while bias might be controlled by the windowing approach in resampling. (6) The methodology presents an option for the resampling of spatially correlated data and of other correlation structures, with practical application in fields such as computer vision, geographic medicine, agriculture, environmental sciences and mining.