Research On The Segmentation And Classification Of Ultrasound Breast Tumor Images Based On Convolution Neural Network

Breast cancer is considered the most common cancer.Ultrasonography is an important clinical diagnostic method for locating breast tumors.Due to ultrasound artifacts,low contrast and complex tumor shapes in ultrasound images,there may be misdiagnosis and missed diagnosis in the process of manual diagnosis.Therefore,the segmentation and classification of breast tumors by automatic ultrasound using computer aided system can help doctors locate and judge the location and type of tumors.It is also improving the diagnostic accuracy and alleviating doctors’ workload.At present,the methods of tumor image recognition based on ultrasound are limited and the accuracy is not high.Therefore,in this paper,an end-to-end automatic tumor segmentation and classification method is proposed by using convolution neural networks for two common data sets.The main innovations of this research are as follows:(1)Aiming at the problems of low quality of ultrasonic images and easy interference of ultrasonic artifact segmentation results,a boundary oriented ultrasonic breast tumor segmentation network was proposed.This network improves the performance of tumor segmentation from two perspectives.First,a boundary oriented module is designed to capture weak boundaries of breast tumors by learning additional breast tumor boundary maps.Secondly,the network is more focused on enhanced feature extraction,using the pooled space pyramid pool module and extrusion excitation block to obtain multi-scale and efficient feature information.Finally,k-fold cross-training was used to alleviate the problem of insufficient data sets.Based on experimental data and results analysis,the proposed boundary-oriented tumor segmentation network has the highest accuracy and stability.(2)The convolution neural network using conventional convolution operation has a large number of parameters and requires a large amount of calculation,which is not conducive to its use in mobile and embedded devices,this paper adopts a lightweight convolution neural network with depth separable convolution operation to classify tumor categories in ultrasonic images.Firstly,the convolution block attention module is introduced into the network to improve the ability of extracting important features from channel and spatial attention.Secondly,focus loss function and label distribution sensing margin loss function were combined to solve the problem of uneven classification of ultrasound breast tumor data sets.Finally,transfer learning is used to alleviate the problems of slow convergence and insufficient robustness caused by insufficient data sets.The experimental results also prove that the lightweight network in this paper has fewer parameters and higher classification accuracy.