Semantic Segmentation Method For Breast Ultrasound Images Based On Fully Convolutional Network

As breast cancer has become one of the most lethal cancers,millions of women suffer every year.Survival rate could be significantly improved when breast cancer can be found early and people get effective treatment.Nowadays,despite various methods for breast cancer detection,breast ultrasound detection technology is adopted widely because of low cost and high cost performance.With the development of computer technology,computer aided diagnosis(CAD)has become an important method to assist doctors in diagnosis.It plays an important role for doctors in strengthening the diagnostic accuracy and objectivity,avoiding misdiagnosis and improving the efficiency.The current CAD system of breast ultrasound segments the tumor region of breast ultrasound by possessing such images,and there are mainly two types: semiautomatic and automatic segmentation.Most common used,semi-automatic method completes segmentation by manually selecting seed points or marking regions of interest,leading to an increase in doctors' burden and misdiagnosis and thus difficulties in promotion.The automatic segmentation method barely requires manual participation,but it can only segment the tumor area and thus ignore other breast tissues' assistance to the doctors' diagnosis.In practical diagnosis,the skin layer,the fat layer,the gland layer and the muscle layer in the breast will serve as an auxiliary reference while the tumor's region,texture and shape serve as the main reference.As a result,the current automatic segmentation methods do not consider the impact of other organizations on the diagnosis of doctors from an overall perspective,which limits segmentation ability.For the above discussed problems,this paper combines the achievements of the convolution neural network in image semantic segmentation and proposes a multi target semantic segmentation method based on full convolution network.This method can automatically divide the corresponding tissues in the breast ultrasound images.The main work is focused on the following two aspects:(1)This paper chooses the Alex Net network structure as the basic network structure of the full convolution network,because Alex Net has low depth,simple structure,and excellent performance in the case of insufficient samples.Next,in view of the deficiency of breast ultrasound samples and the characteristics of ultrasonic images,this paper proposes enhancing the image by the method of wavelet transform and expanding the data set by rotating and turning.The network training combines the characteristics of the whole convolution network and those of the data set,and uses the same training parameters to train and segment the different regions separately.At last,the semantic segmentation of the breast ultrasound image is completed by integrating the segmentation results of different organizations.(2)The idea of jumping structure in the convolution network is used to improve the Alex Net structure and further to increase the segmentation accuracy by fully connected conditional random field.Due to the influence of convolution and pooling in CNN,the segmentation results of the FCN model,before the improvement,are relatively rough,the detail expression of the specific parts is not specific enough,and there are many areas of error segmentation.In this paper,the idea of jumping structure is used to fuse the model output and output of the local normalized response layer in the Alex Net network,and thus the loss caused by the convolution and pooling is decreased.And also the same training strategy is used to improve the accuracy of the segmentation.Next,considering the lack of spatial consistency and ignorance of the relationship between the pixels in the convolutional network,this paper uses the fully connected conditional random field to optimize.The Fully-connected conditional random filed takes the probability graph output of the improved model as the one element energy term,uses the pixel value of the original image and the pixel position to construct the two element energy term,and obtain the optimal segmentation result by solving the energy minimization through the mean field approximation method.