Computer aided detection of spina bifida using nearest neighbor classification with curvature scale space features of fetal skulls extracted from ultrasound images

This paper addresses the problem of detecting the common neural tube defect of spina bifida by a computer aided detection (CAD) system. We propose a method which extracts the curvature scale space (CSS) features of fetal skull contours viewed in the ultrasound (US) modality and performs nearest neighbor (kNN) classification on those features having the desired properties of invariance with respect to translation, orientation and scale changes, thus improving robustness. The distance between two sets of CSS features, each set corresponding to the description of the contour of a particular skull, is measured as the cost of matching the two sets of CSS features. Such a CAD system may act as a second observer and help experts in prenatal diagnosis.Our data possess absolute and relative rarity. The experiments are performed with two different rare class handling methods and over a range of operating conditions. All experiments are based on a group of settings; associated with using either balanced or unbalanced datasets, employing different types of CSS features and how CSS matching costs are computed. Comparatively evaluating the classification performance of the settings is carried with the aid of the whole-curve metric of area under the receiver operating characteristics (ROC) curve (AUC). Optimal operating conditions for any setting can be identified and some settings reveal advantages over others. The observations indicate that using balanced datasets offers better performance and our proposed version of estimating CSS matching costs is generally superior to the classical method. Furthermore, using enhanced sets of CSS features improves classification accuracy. When classification is performed on balanced data using enhanced CSS features and the matching cost is computed with our proposed technique; one can observe an F-measure of 0.76 along with 70% TP rate (recall), 17% FP rate (false alarms) and 82% precision.