Automated detection of multiple sclerosis candidate regions in MR images: False-positive removal with use of an ANN-controlled level-set method

Our purpose in this study was to develop an automated segmentation scheme for multiple sclerosis (MS) lesions in magnetic resonance images using an artificial neural network (ANN)-controlled level-set method. Forty-nine slices with T1-weighted, T2-weighted, and fluid-attenuated inversion recovery images were selected from six examinations of three MS patients including 168 MS lesions for this study. First, MS lesions were enhanced by background subtraction. Initial regions of MS candidates were detected based on a multiple-gray-level thresholding technique and a region-growing technique on the subtraction image. Then, final regions of MS candidates were determined by application of a proposed segmentation method using an ANN-controlled level-set method, which was used for reduction of false positives (FPs) as well as more accurate segmentation. Finally, all candidate regions were classified into true positive and FP candidate regions by use of a support vector machine. As the result of a leave-one-candidate-out test method, the detection sensitivity for MS lesions increased from 64.9 to 75.0% while decreasing the number of FPs per slice from 19.9 to 4.4 compared with a previous study. The proposed scheme improved the sensitivity and the number of FPs in the detection of MS lesions.