Breast Ultrasound Image Classification On Deep Feature Based Transfer Learning And Feature Fusion

Breast cancer is the second leading cause of death for women all over the world.The cause of the disease has been unknown so far and early detection and diagnosis is the most effective way,which can increase the success rate of treatment,save lives and reduce costs.Ultrasound imaging is one of the most commonly used diagnostic tools for early detection and classification of breast cancer,with its advantages of easy operation,cost-effectiveness,non-invasive and real-time diagnosis capabilities.Generally,radiologists read and judge breast ultrasound images with strong subjectivity while the clinical application of computer-aided diagnosis(CAD)system avoids the shortcoming for increasing the accuracy,objectivity and repeatability of breast masses detection and classification as a valuable and beneficial means.Currently,the rapid development of computer-aided diagnosis technology provides much assistance to diagnosis of breast disease,however the level of promotion and application is far from enough due to its limitations and deficiencies.The emerging deep learning method has much potential in the research direction,although its application and progress in breast ultrasound image processing and recognition are far less than natural image.This research aims to address the problem of discriminating benign cysts from malignant masses in breast ultrasound images based on deep convolutional features.Breast ultrasound(BUS)image dataset in this research was acquired directly from 1422 patients in The Third Affiliated Hospital of Sun Yat-sen University,which contains 2058 cases with 688 malignant solid masses and 1370 benign masses.All the diagnosis results of the cases were confirmed by both biopsy and operation with high credibility.This thesis firstly reviewed the present situation of breast ultrasound image detection and classification.After that,we studied the framework algorithms of classification based on traditional feature extraction and deep convolutional feature extraction respectively.At last,the classification results were compared and analyzed.Main contributions of this thesis are:(1)Study on breast masses classification based on traditional feature extraction.First,analyze breast ultrasound image including image de-noising preprocessing,tumor segmentation.Then extract multiple discriminative features and introduce efficient classifiers for breast masses classification.Specifically,first-order features,texture features and shape features were extracted from the breast ultrasound images.The effectiveness of the classification method based on traditional feature extraction was compared and evaluated by using two classification algorithms like SVM and Adaboost.(2)Research and analysis on breast masses classification performance based on deep convolutional features.Based on BUS dataset,a specific convolutional neural network structure was designed for automatic feature extraction of breast ultrasound images,and different types of image features were obtained for classification.The aim of this experiment is to introduce a new framework algorithm of classification for breast ultrasound image,and to achieve better classification performance compared with the traditional feature extraction classification method.(3)Research and analysis on breast masses classification based on feature based transfer learning.Deep feature based transfer learning was applied to extract features and classify breast ultrasound images with the consideration of lacking image samples,which utilizes the pre-trained deep neural network on large-scale natural image dataset and fine-tunes it on our own BUS image dataset.The model was evaluated by cross-validation method,and a variety of evaluation indexes were adopted to analyze the effect of classification.Furthermore,different transferred deep neural networks were trained and compared in differentiating benign and malignant tumors on the BUS dataset.(4)Fusing the deep convolutional features extracted from transferred deep neural networks,and making it as the input feature vector of Artificial Neural Network(ANN).The effectiveness of the method was verified and analyzed on the BUS dataset for the differentiation and classification of breast masses.The research work of this thesis indicates that a classification method based on deep convolutional features can improve the accuracy of breast ultrasound image classification to some extent.The model not only can effectively avoid the tedious steps of traditional manually designing features,but also significantly improve classification accuracy.Additionally,the classification performance is further improved by using feature based transfer learning,which alleviates the effect of lack of image data on breast ultrasound image processing based on deep convolutional feature.Last but not least,the fusion of features extracted from top performing deep neural networks using feature based transfer learning further boosts the classification performance.