Automatic Boundary Detection in Breast Ultrasound Images Based on Improved Pulse Coupled Neural Network and Active Contour Model

Due to the speckle noise, low signal/noise ratio, low contrast and blurry boundaries in breast ultrasound (BUS) images, fully automatic boundary detection from the BUS image is a difficult task. To solve this problem, a novel segmentation method for the BUS image is proposed by using a two-stage strategy: the region of interest (ROI) generation and segmentation. First, an improved pulse coupled neural network (PCNN) model is used to categorize the image into different classes from which the ROI is selected by using the background rules. The rough contour for the breast tumor is also obtained in this stage. Then, the rough contour is used as the initial condition for the modified active contours without edges (ACWE) to get the final boundary of the breast tumor. To demonstrate the effectiveness of this algorithm, experiments are designed to compare the algorithm results with the corresponding tumor regions delineated by an experienced radiologist. Three error metrics, true-positive ratio (TP), false-negative ratio (FN) and false-positive ratio (FP) are used to measure the segmentation performance. The final results with TP= 96.7%, FN=3.12%, and FP=5.50% demonstrate that the proposed method can segment BUS images efficiently and automatically.