Research On Breast Mass Detection And Classification Based On Deep Learnin

With the accelerating pace of people’s lives,breast cancer has become the most common cancer at present,which is seriously endangering the health of women.Early detection and treatment of breast cancer play an important role in improving the cure rate.The most commonly used method of breast cancer is mammography.As a research object in this work,mammography is combined with computer-aided diagnosis technology to achieve automatic detection and classification of breast masses,which can greatly help doctors to make more rapid and accurate diagnosis of the mass.The research in this paper has far-reaching significance for early diagnosis and treatment of breast cancer.It is helpful for doctors to judge the mass more quickly and accurately.In order to complete the two tasks of mass detection and classification,the following works are done in this work:(1)Breast mass detection algorithm based on improved YOLOv4In order to detect the location of the mass more quickly and accurately,an improved algorithm based on YOLOv4 is proposed in this work.Firstly,in order to improve the detection efficiency of network,the backbone network of YOLOv4 algorithm is replaced by MobileNetv3 network,which makes the network lighter.Secondly,in order to improve the accuracy of network recognition,the initial anchor is redesigned by K-means++ algorithm to make it more in line with the size of breast mass,and excellent improve the accuracy of network recognition.Finally,in order to improve the detection ability of the network,the number of convolutions in the network is increased and the depthwise separable convolution is simultaneously introduced in this work,greatly reduces the training parameters of the network without affecting the feature extraction.Experiments show that comparing with the original network and other mainstream networks,the proposed algorithm has better detection accuracy and speed.(2)Breast mass classification algorithm based on EfficientDetAccurate judgment of the type of mass is helpful for doctors to carry out the next step of treatment.In this work,the masses are devided into BI-RADS 2,3,4,5.Due to the subdivision of tumor categories,the feature differences between different types of tumors are small,which increases the difficulty of classification.Therefore,this paper applies EfficientDet to the task of breast tumor classification for the first time,and concurrently the attention mechanism of CBAM spatial channel mixing is introduced to improve the learning ability of the network to key areas.In the aspect of training strategy,the learning rate decline mode is improved to cosine annealing decline mode to heighten the ability of network training.The sensitivity of the final improved model is 79.86%,and the mAP is 82.73%.