Research On Saliency Detection For Breast Lesions Based On Multiple Prior Knowledge

Breast cancer is one of the deadly cancers for women.Clinical studies have found that early detection of breast lesions is conducive to improving the success rate of cure.Therefore,early diagnosis methods are essential.Ultrasound imaging technology has become one of the most popular early breast cancer screening methods due to its non-invasiveness,safety and high sensitivity,and its ability to show the changes in the morphology,internal structure and adjacent tissues of each layer of masses.Breast ultrasound image analysis is usually one of the strategies used by radiologists to provide correct diagnosis for patients.The size,shape and boundary of breast lesions are crucial for identifying the type of lesion.Therefore,clinical experience and professional knowledge is usually significant for diagnosis.In order to assist doctors in obtaining accurate and reliable diagnosis,many computer-aided diagnosis(CAD)systems have been established in recent years.The automatic detection of breast lesions is one of the most important components of the CAD system.Although many breast lesion detection methods have been proposed,they still have some shortcomings.For example,low-contrast breast ultrasound images are usually processed with low precision and poor performance;since breast lesion detection is incomplete,some regions of lesions most medically concerned cannot be well detected.To deal with these problems,this thesis proposes a saliency detection model for breast lesions,which incorporates background prior,frequency prior,and adaptive center prior.The model combines the background prior information,frequency prior information,and center prior information of breast ultrasound images to calculate the saliency value of breast lesions,and then optimizes the detection results by using single cell automata and Graph-cut to achieve the final breast lesion saliency detection.Finally,by combining the original and detected breast ultrasound images,the boundary heat map of breast lesions is introduced to visually reveal the difference between benign and malignant lesions.The main research contents are as follows.(1)To deal with the insufficient prior knowledge of breast lesions in the previous saliency detection models,a saliency detection model for breast lesions is proposed based on multiple prior saliency maps.The model first uses some image preprocessing methods to enhance breast ultrasound images.Then,considering the relationship between the lesion and the background of the breast ultrasound image,the model selects the image boundary as the background seed to calculate the saliency of the lesion and generates a rough saliency map of the breast lesion based on the background prior.Because the obtained rough saliency map still has some problems of missed detection and misdetection of the lesions,it is further optimized by combining the frequency prior and the adaptive center prior information.Experiments show that the breast lesion saliency detection method,which combines multiple prior knowledge,can accurately locate the breast lesion area,and is better than the saliency detection methods with single prior knowledge.(2)Aiming at the problem that the proposed saliency detection model based on multiple prior saliency maps cannot meet the clinical high-precision requirements,an optimization method is porposed based on the combination of single cell automata and Graph-cut.This method introduces an impact factor matrix and a coherence matrix to iteratively update the rough saliency map based on the background prior,and further optimizes the final breast lesion saliency map by combining the Graph-cut technology.Experiments show that the model with the optimization method can effectively eliminate a large number of non-lesion areas and accurately locate breast lesion areas.(3)In view of the previous breast lesion detection methods lacking of analyzing the types of breast lesions through the lesion boundary,this thesis presents a method for analyzing the types of breast lesions based on the boundary heat map of breast lesions.This method visualizes the breast lesion by superimposing the saliency map of the breast lesion on the original breast ultrasound image to generate the lesion boundary heat map,which explains the difference in the detection results of benign and malignant breast lesions,and has potential assistance in improving the diagnosis efficiency of doctors effect.