Breast Ultrasound Segmentation Using Evolutionary Pulse-coupled Neural Networks

In this article we present a segmentation algorithm based on computational intelligence paradigms for breast lesions on ultrasound. The parameter tuning of three different pulse-coupled neural networks (PCNN) models was performed through two distinct variants of differential evolution (DE) approach. To demonstrate the effectiveness of these hybrid models (i.e. evolutionary PCNN), a set of experiments was designed to compare the computerized segmentation outcomes with 51 breast tumors delineated by a senior radiologist. Since the proposed algorithm has a stochastic basis, 31 runs of the segmentation method were performed for every image in dataset. The segmentation performance was assessed in terms of accuracy (mean±standard deviation), which was computed from four metrics of area error: true positive, true negative, false positive, and false negative. Also, the percentage of outliers (i.e. atypical low accuracy values) was measured. The results pointed out that the simplified PCNN optimized by DE/rand/1/exp variant attached the best accuracy, 97.06±1.38 %, and the lowest percentage of outliers, 1.62 %.