Design And Implementation Of Breast Tumor Color Ultrasound Detection System Based On Deep Auto-Encoder

Breast tumors are extremely common and frequent among breast surgical diseases and have a high mortality rate when they become malignant,which seriously endangers the physical and mental health of patients.Therefore,early screening to prevent breast tumor deterioration is particularly important,and early screening with ultrasound images is an effective and inexpensive means.However,with the increasing number of patients,the pressure on radiologists to identify images is gradually increasing,and there are also differences in experience among doctors,leading to an increase in the rate of misdiagnosis and omission.To alleviate the pressure of doctors’ image recognition,building deep learning models for breast image assisted detection has gradually become a research hotspot for domestic and international scholars.But there are still some problems in the existing studies: due to the differences in the characteristics of Chinese and foreign populations,it is inappropriate to use models trained with foreign datasets for direct application to Chinese patients;the datasets of related studies for domestic patients are generally small,which easily leads to insufficient model capability;the number of benign and malignant data used in the related studies is extremely unbalanced,and more benign data easily leads to data waste and is difficult to be utilized rationally;In ultrasound images of breast tumors,the features that determine the nature of tumors have different levels of abstraction,and the traditional convolutional neural network cannot well take into account the image regions of different sizes,making it difficult to fit the features of breast tumor ultrasound images.To address the above problems,the following work is carried out in this thesis:First,the datasets collected in this thesis are all real breast ultrasound datasets from a hospital in southwest China,so the model trained with Chinese population data can be more specific to the unique characteristics of the Chinese population and avoid the model judgment errors caused by the differences in ultrasound imaging caused by the differences in human body characteristics between domestic and foreign populations.Second,this thesis uses a relatively large number of 7188 images in the breast tumor ultrasound dataset,and uses transfer learning to fine-tune the model to further prevent the performance shortage caused by the data volume.At the same time,the deep auto-encoder idea is used to construct the model,which not only makes reasonable and effective use of a large amount of benign data,but also allows the network to further improve the feature extraction ability of the network based on transfer learning.In addition,the use of Focal Loss further alleviates the problems caused by data imbalance.Then,based on the above transfer learning and auto-encoder ideas,the Dense Net121-Autoencoder model was constructed based on the Dense Net network.The densely connected structure of Dense Net enhances the propagation of features and transfers the shallow features to the deep layer,which enables the image features of different levels of abstraction to jointly determine the final result and better fits the characteristics of breast tumor ultrasound images.The final Dense Net121-Autoencoder model achieves an accuracy of 93.98% on the internal test set,which is the same as the accuracy without the auto-encoder structure while improving the recall rate and 0.75 percentage points higher than the accuracy without transfer learning.At the same time,the Dense Net-based network obtained better performance results compared to the same parameter volume models based on VGG16,VGG19,Res Net50,and Mobile Net V2,all using the auto-encoder structure.Finally,a breast tumor ultrasound image detection system is designed and implemented using Think PHP 6.0 backend framework and Layui front-end development tool,integrating a pre-trained Dense Net121-Autoencoder model to support the determination of benign and malignant breast ultrasound images.At the same time,this thesis implements the breast tumor ultrasound image detection system to support user login,role management,image enhancement and other functions,so as to achieve the goal of systematically assisting doctors in diagnosis.