Breast Ultrasound Tumor Detection Based On Convolutional Neural Network

Breast cancer is the most common cancer among women.If breast cancer can be detected and treated early,the survival rate of patients will be greatly improved.Ultrasound imaging is widely adopted due to its low radioactivity,good detection effect,and low price.Compared with other types of images,breast ultrasound tumor images have lower resolution,greater noise,and more blurred boundaries.Relying only on manual diagnosis will increase the misdiagnos greatly.The recognition methods based on machine learning often have disadvantages of low accuracy and poor versatility.At present,most of the recognition methods based on deep learning are limited to locate the target area or classify the given target area as benign or malignant,which bring great inconvenience to medical staff.Although the latest research introduced Faster R-CNN algorithm,which realized the classification and location of tumors at the same time,however,the algorithm have problems of low accuracy and slow speed.Given this,this article continues to follow this idea that using target detection algorithms based on deep learning to achieve tumor recognition and improve the algorithm for related problems in tumor recognition.The main research contents of this article include the following:1.The four classic target detection algorithms based on deep learning(Faster R-CNN,SSD,YOLOV3 and CornerNet)are applied to the identification of breast tumor ultrasound tumors and the algorithm with higher recognition accuracy and rate is screened out.First of all,this article analyzes the principles of the four algorithms;Then,the four algorithms are trained and tested on the pre-trained model called Image Net with the help of the data set marked by professional doctors;Finally,various evaluation indicators such as mAP and FPS are used to evaluate the performance of four algorithms.Experimental results show that YOLOV3 has achieved better results in accuracy and speed compared with the other three algorithms.2.Aiming at the problem that the features of breast ultrasound tumor are more difficult to extract than other types of images,DarkNet-53 is improved,which is the feature extraction network of YOLOV3 algorithm.First of all,the SE module and Res2Net network are merged into the SE-Res2Net network to expand the receptive field in a single layer and extract more full convolution features by replacing the original Res Net;Then on the basis of the above,aiming at the problem that the 3 ~* 3 convolutions used in downsampling is easy to cause the defect of feature loss,three different Downsample of downsampling modules are proposed;Next,on the basis of the above,residual connection and dense connection are introduced to form a new network called Res-DenseNet to improve the original residual connection method to further improve the ability of feature extraction;Finally the algorithms are trained and tested separately.Experiments show that these three improvements can improve the recognition effect of YOLOV3 algorithm.3.On the basis of the previous algorithm,YOLOV3 algorithm continues to be improved in response to the problem of large number of small targets in breast ultrasound tumor images.Frist of all,the three-scale prediction network in the YOLOV3 algorithm is improved to four scales to make full use of the location information of the shallow feature map,and considering the problem of the parameter amount,the network is cropped;Secondly,the feature fusion method of the original algorithm called FPN is replaced by a bidirectional pyramid network called ReBiF to fully integrate location information of the shallow network and semantic information of deep network;Finally,the above improved algorithm is trained and tested separately.The experimental results show that on the basis of the foregoing,the accuracy of the YOLOV3 algorithm has been greatly increased after improving the algorithm,and its detection rate has decreased compared to the original algorithm,but it is still higher than the SSD algorithm.