Background Segmentation Of Ultrasound CT Images Based On Deep Learning

According to the latest statistics released in 2021,breast cancer has the highest incidence among all diagnosed cancer patients,and it has gradually become one of the main causes of death.Clinically,the use of ultrasound CT for breast cancer examinations for patients is expected to become a new and reliable examination method for breast cancer screening due to its advantages of no radiation,strong real-time,no squeezing and high sensitivity to dense breasts.Because the collection cavity contains water and air bubbles,there may be a lot of background noise in the ultrasound CT image,which will interfere and affect the threedimensional imaging and subsequent clinical diagnosis.Therefore,it is necessary to separate the inspected area and the background area in the image,but the manual segmentation is time-consuming and laborious,which increases the burden on the doctor.Therefore,the study of a reliable automatic segmentation algorithm for ultrasound CT images has important application value.Deep Learning method currently performs prominently in image segmentation tasks and is a very popular segmentation algorithm.It is very feasible to study the use of the powerful feature extraction capabilities of deep learning for background segmentation of ultrasound CT images,and it is of great significance for the subsequent Computer Aided Diagnosis(CAD).In this thesis,the collected original ultrasound CT images are manually annotated to construct a training data set.Including ultrasound CT images of the breasts of 12 volunteers and ultrasound CT images of the arms and palms of 2 researchers.In order to solve the problem of insufficient data set,the use of data enhancement methods can effectively expand the sample size of the data set.Three convolutional neural networks: UNet,UNet++ and CBAM-UNet are used for background segmentation experiments of ultrasound CT images.CBAM-UNet is an improvement on UNet,which combines convolutional block attention module(CBAM)composed of channel attention and spatial attention.In addition,this paper also proposes Transformer UNet using the popular Transformer structure in the field of natural language processing to replace the traditional convolutional down-sampling.By comparing the results of image segmentation and calculating quantitative evaluation indicators,the results showed that among the three convolutional neural networks,UNet++has the best segmentation performance,but the training time is longer;CBAM-UNet is superior to UNet in subjective evaluation of segmentation results,but the objective quantitative evaluation index is slightly worse.The segmentation results of the Transformer UNet network are better than these three convolutional neural networks,which proves that the Transformer UNet model proposed in this paper can better extract image features in image segmentation tasks,and has superior transfer learning capabilities.