A Study On Ultrasound Image Classification Of Breast Tumors Based On Migration Learning

Breast cancer is the most common malignant tumor in modern women,and early screening is a very important means of preventing breast cancer.In China,due to the imbalance of urban and rural development,large population base,and differences in medical standards,female breast cancer screening at the grassroots level lags behind cities.Beginning in 2009,"Two Cancer Screening" was included in the government work report,which plans to screen 1.2 million rural women for breast cancer every year.Among the early screening methods for breast cancer,ultrasonography is widely used in the grassroots because of its high cost performance,no radiation,and few side effects.However,ultrasound examination relies heavily on doctors’reading and interpretation of superficial images.It takes a long time and a large workload,and it is difficult to meet the needs of rapid and batch clinical diagnosis.In order to improve the diagnostic efficiency and reduce the labor intensity of doctors,a computer-aided diagnosis system for breast ultrasound images has emerged.This paper studies the transfer model based on depth model to classify ultrasound images of breast tumors.The experiment firstly preprocessed the ultrasound images of breast tumors,and then carried out pre-training and transfer training of the model.Finally,the classification performance of the model was evaluated.The transfer model uses the Google Inception v3 model,which excels in the field of deep learning classification.By reserving the convolutional layer parameters of the classification model,re-optimizing the fully connected layer structure and retraining the full connection layer parameters with breast tumor ultrasound images to achieve classification.the goal of.Model transfer training uses the Dropout algorithm to reduce overfitting,and the Adam algorithm is used to optimize the parameters of the network model.The re-optimized breast ultrasound image classification model in the logits layer uses a combination of SVM classifiers and sigmoid activation functions that are superior in the two classifications.The ultrasound data set of breast tumors was divided into training set,verification set and test set.The classification and results of the verification set were stable at 80%,and the real-time test results in the model training process reached 94.1%.After the model was exported,the performance of the model was tested using 100 ultrasound data that did not participate in the training.The classification accuracy rate reached 85%.