Fully Automatic Segmentation For Breast Ultrasound Image Based On Saliency Detection

Breast cancer is a common malignant tumor and one of the most often diagnosed in women. It seriously endangers both mental and physical health of women. Early detection, diagnosis and treatment of cancer is the best way for fighting cancer recognized by the international medical community. Ultrasound with high precision, low cost. It has no trauma to the human body, no radiation, and other advantages. It becomes the main imaging means for the diagnosis of breast cancer. The application of computer aided diagnosis(CAD) could assist doctors to judge, improve the objectivity and accuracy of the results, reduce the misdiagnosis and missed diagnosis, and is important for the diagnosis and treatment of disease.The segmentation of tumor is a key and challenging step in the breast ultrasound CAD system. Most of current breast tumor segmentation methods are semi-automatic. In the process of the algorithms, these methods need manual intervention or guidance such as drawing the region of interest, selecting seed point, etc. When faced with a large number of tasks, manual operation becomes the bottleneck. In interpretation of breast ultrasound image, doctor would modify the result in a single image based on the relationship among the image sequences to get more accurate conclusions. Most of current breast tumor segmentation methods take single ultrasonic image as the input unit. And there is a lack of global consistency in the results of segmentation.Aiming at the problem existing in the breast tumor segmentation methods, fully automatic segmentation for breast ultrasound image based on saliency detection is proposed in this paper. The main contents include the following two aspects:First of all, this paper proposes a breast tumor segmentation method based on single image saliency detection. The method constructs location prior saliency map and background cue saliency map based on single ultrasound image, which combined with the anatomical structure of the breast and the characteristics of ultrasound image. The two saliency maps have their own advantages in reflecting the saliency of the breast tumor. The initial saliency map is generated by the fusion of these two saliency maps. Aiming at the problem that uneven and non-compact interior existed in the salient object(tumor) in the initial saliency map, the final saliency map is obtained by saliency value propagation method which combined with fuzzy connectedness. Experimental results show that the saliency map generated by the proposed method can eliminate the background effectively and highlight the tumor region.Secondly, this paper proposes a breast tumor segmentation method based on video saliency detection and random walks algorithm. The method takes an image sequence as processing unit, constructs a graph model with superpixels and defines the inner neighbors and the inter neighbors of nodes to reflect the local information of the image regions. Updating the saliency map based on single image information by the synchronous saliency value updating method which contains the neighbor relationship of nodes to obtain the results which are more uniform, more compact and with high consistency. On the basis of video saliency detection, we generate a large amount of the foreground and background seed points which are uniformly distributed according to the saliency value, and combine with the random walks algorithm to generate the detection results with stronger expression ability at the pixel level. Experimental results show that the saliency values of the saliency map calculated by the method are more uniform and compact. A clear boundary can be formed between the tumor and the background and the results with high consistency in the image sequence.