Interactive breast mass segmentation using a convex active contour model with optimal threshold values

Abstract

IntroductionA convex active contour model requires a predefined threshold value to determine the global solution for the best contour to use when doing mass segmentation. Fixed thresholds or manual tuning of threshold values for optimum mass boundary delineation are impracticable.A proposed method is presented to determine an optimized mass-specific threshold value for the convex active contour derived from the probability matrix of the mass with the particle swarm optimization method. We compared our results with the Chan–Vese segmentation and a published global segmentation model on masses detected on direct digital mammograms. Methods and materialsThe regional term of the convex active contour model maximizes the posterior partitioning probability for binary segmentation. Suppose the probability matrix is binary thresholded using the particle swarm optimization to obtain a value T1, we define the optimal threshold value for the global minimizer of the convex active contour as the mean intensity of all pixels whose probabilities are greater than T1. ResultsThe mean Jaccard similarity indices were 0.89±0.07 for the proposed/Chan–Vese method and 0.88±0.06 for the proposed/published segmentation model. The mean Euclidean distance between Fourier descriptors of the segmented areas was 0.05±0.03 for the proposed/Chan–Vese method and 0.06±0.04 for the proposed/published segmentation model. ConclusionsThis efficient method avoids problems of initial level set contour placement and contour re-initialization. Moreover, optimum segmentation results are realized for all masses improving on the fixed threshold value of 0.5 proposed elsewhere.