Convolutional Neural Networks based classification of breast ultrasonography images by hybrid method with respect to benign, malignant, and normal using mRMR

Highlights

• Breast cancer ultrasound images were diagnosed and classified with the proposed hybrid model.

• In this study, features obtained from Alexnet, Mobilenet, and Resnet50 architectures were concatenated. (It has been observed that the hybrid feature extraction structure yielded better results).

• The accuracy rate of the hybrid model is 95.6%.

Abstract

Early diagnosis of breast lesions and differentiation of malignant lesions from benign lesions are important for the prognosis of breast cancer. In the diagnosis of this disease ultrasound is an extremely important radiological imaging method because it enables biopsy as well as lesion characterization. Since ultrasonographic diagnosis depends on the expert, the knowledge level and experience of the user is very important. In addition, the contribution of computer aided systems is quite high, as these systems can reduce the workload of radiologists and reinforce their knowledge and experience when considered together with a dense patient population in hospital conditions. In this paper, a hybrid based CNN system is developed for diagnosing breast cancer lesions with respect to benign, malignant and normal. Alexnet, MobilenetV2, and Resnet50 models are used as the base for the Hybrid structure. The features of these models used are obtained and concatenated separately. Thus, the number of features used are increased. Later, the most valuable of these features are selected by the mRMR (Minimum Redundancy Maximum Relevance) feature selection method and classified with machine learning classifiers such as SVM, KNN. The highest rate is obtained in the SVM classifier with 95.6% in accuracy.

Introduction

Breast lesions are three-dimensional space-occupying lesions of the breast. Benign and malignant breast lesions consist of many diseases with very different radiological features. Breast cancer is the most common type of cancer in women and ranks second in cancer-related deaths after lung cancer [1]. Therefore, early diagnosis of breast lesions and differentiation of malignant lesions from benign lesions are extremely important for the prognosis of patients [2]. The main radiological methods used in imaging breast lesions are mammography, ultrasonography, and magnetic resonance imaging. Mammography is soft tissue radiography that allows the breast to be viewed. Breast ultrasonography is a nonionizing and noninvasive imaging method in which sound waves are used. It is an extremely important imaging method in distinguishing solid-cystic breast lesions and characterizing solid lesions. It provides important information about the vascularity of the lesions with Doppler ultrasonography. It also guides the biopsy of breast lesions. Recently, with the use of high-frequency ultrasonography probes with advanced technologies, even small breast lesions can be seen in ultrasonography. In addition, since the mammography of patients with dense breasts is not sufficiently diagnostic, ultrasonography is used in these patients in order not to overlook the lesion and to diagnose it at an early stage [3]. Ultrasonography is preferred especially in young patients to prevent radiation exposure of the breast. In addition to being inexpensive, easily accessible, reproducible, and not containing ionizing radiation, breast ultrasonography has important advantages [4]. Our aim in this study is to classify benign, malignant, and normal breast lesions by using deep learning methods in ultrasonographic images. With the developing technology, deep learning architectures are becoming popular day by day [5].

In this study, three different CNN models such as Alexnet, Resnet50, and MobilenetV2 are used for concatenating features. Later, the features taken from these architectures are classified with various classifiers. Finally, in the proposed model, features from Alexnet, Resnet50, and MobilenetV2 are concatenated. The mRMR feature selection method is used to reduce the size of these concatenated features and to save time and cost. The best features obtained with mRMR have been classified in various classifiers such as SVM, KNN which are very popular machine learning classifiers. According to the results we obtained in this study, it shows that computer-aided systems can be used more frequently in medical application [6].

There are articles on breast cancer classification in the literature. Kwon et al. proposed an automatic breast ultrasound system for breast cancer evaluation. Here, two different imaging techniques have been used to compare their diagnostic performances. 136 asymptomatic women were used in the study. It was observed that the sensitivity and specifity values were similar in both imaging techniques. In this study, the specifity values of 97.7% in the 2-VST method and 95.2% in the 3-VST method were obtained [7].

Sun et al. conducted a classification procedure using 2395 breast ultrasound images of 479 breast cancer patients. In the CNN model they used, they worked with two classes of images, benign and malignant. His AUC value in his studies is 95%. They also stated that the CNN model they used was Densenet201 [8].

Zhang et al. stated that they collected ultrasound images from 2 hospitals in their study. They stated that there were 2822, 707, 210 images, respectively, in the data set they obtained, and these images were used both in training the model and testing the model. In this study, the accuracy value they obtained was 92.86. They used deep learning architectures in their study [9].

Kim et al. focused on the ABUS method in their work on automatic breast ultrasound scanning for dense breasts. They shared the use of this method and the results they obtained in ultrasound images. It was seen that different sensitivity and specifity values were obtained here [10].

Masud et al. classified breast cancer ultrasound images in their study. Here they combined 2 different data sets. They used pre-trained CNN models in their work. They divided the data in the data set into 3 classes with the CNN model they proposed. These classes are malignant, benign, and normal. In the model they recommend, the accuracy value is 100% and the AUC value is 1 [11].

In their study, Zhang et al. analysed four pre-trained CNN models in a dataset of ultrasound images. They achieved the highest AUC of 0.90 among these pre-trained models [12].

Moon et al. aimed to diagnose breast cancer using ultrasound images in their study. They used 2 different data sets in their studies. Using the image fusion method, they brought together different CNN methods. They stated that they achieved an accuracy of 91.10% in their first dataset and 94.62% in their second dataset [13].

Kim et al. Proposed a deep learning-based method using breast ultrasound images in their study. Different performance criteria were calculated in this study. The AUC value was found to be 89% [14].

Moon et al. used ultrasound images for the diagnosis of breast cancer in their study. The difference between the proposed method and the traditional method is that they used 3-D Convolutional Neural Networks architectures. They have calculated different performance metrics in these studies they have done. The accuracy rate they achieved in their study was 96.7% [15].

Breast cancer ultrasound images are classified and presented to the expert in this study. It will be easier for the expert to interpret these images and individual mistakes will be avoided. In addition, thanks to computer-aided systems, the workload of experts will be reduced. This system offers the possibility to use it for pre-diagnosis in non-specialist places. The roughly hybrid model is presented in Fig. 1.

Breast cancer is getting more common day by day. Interpretation of each patient's data by the specialist can cause problems in terms of time and cost. Computer-aided systems can be used in the diagnosis process to prevent errors that may occur in traditional methods. A CNN-based hybrid model has been proposed to diagnose and classify breast cancer images, especially in images. An accuracy rate of 95.6% was achieved with the proposed hybrid model. When the obtained results are compared with the studies in the literature, it has been observed that the proposed hybrid method reaches a higher accuracy value. It has been followed that the developed hybrid model can successfully classify breast cancer ultrasound images.

In the first section of the paper, general information is given, and the relevant literature has been examined. The purpose of the paper is mentioned in the 1.2. contribution. In chapter two, the data set used is examined, then the structure of the developed hybrid architecture is detailed. In chapter three, the results and details of the study are given, the performance criteria used in CNN models are also examined, and the performance criteria of the proposed hybrid model are evaluated. Chapter four is discussion and final chapter is conclusion.