Classification of Breast Ultrasound Images Based on Convolutional Neural Networks - A Comparative Study
Samir M. Badawy, Abd El–Naser A. Mohamed, Alaa A. Hefnawy, Hassan E. Zidan, Mohammed Tarek GadAllah, Ghada M. El‐Banby
Abstract
Computer aided diagnosis (CAD) helps physicians towards an early characterization of tumors in different biomedical tissues, including Breast. Deep learning (DL) based image classification, especially convolutional neural networks (CNNs), has achieved a noticeable success in automatic differentiation of breast ultrasound (BUS) images through the last few years. In this paper: ten well-known pretrained CNNs classification models (ResNet18, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, GoogleNet, MobilenetV2, SqueezeNet, DenseNet201, and Xception) have been utilized to classify BUS images by transfer learning (TL). A dataset of 780 BUS images (133 normal, 437 benign, and 210 malignant) has been utilized in training and validation. A selected 375 BUS images from the dataset have been utilized to evaluate the classification's accuracy of each CNN model after TL process. Ultrasound's speckle noise effect on classification has been studied by applying a simulated multiplicative speckle noise to the mentioned 375 BUS images before being classified by the ten CNN models. Three restoration approaches have been applied to treat speckled BUS images before being classified. Accuracy evaluation results have represented different values for each CNN model over all 375 BUS images through different studied circumstances. The best accuracy's value was for: ResNet101 when the input images were clear with high resolution, SqueezeNet in case of speckled input images, and InceptionResNetV2 in case of restored images using the three applied restoration schemes. So, in case of speckled BUS images, it is recommended to utilize a proper preprocessing step for image enhancement before applying a CNN based classification model (InceptionResNetV2 is preferred). Ten modified and trained CNNs models for BUS images' classification (input image size: 128 by 128 by 3) utilized in this study are available to researchers at: https://www.kaggle.com/mohammedtgadallah/ten-cnns-for-breast-us-images-classification