Convolutional Neural Networks Based Classification of Segmented Breast Ultrasound Images – A Comparative Preliminary Study
Mohammed Tarek GadAllah, Abd El–Naser A. Mohamed, Alaa A. Hefnawy, Hassan E. Zidan, Ghada M. El‐Banby, Samir Mohamed Badawy · 2023
Computer-aided diagnosis (CAD) helps physicians in tumors’ identification into various biomedical tissues, breast is one. Through the last few years, image classification techniques based on deep learning (DL), have obtained noticeable success in differentiating breast ultrasound images automatically. More researches, concerned with breast cancer detection, are representing the output as a black and white segmented image (mask) or a selected region on the breast ultrasound image which will in role submitted to physicians to help them in making a better diagnosis decision. Our paper represents the second step. In this paper: ten pre-trained Convolutional neural networks (CNNs) classification models (ResNet18, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, GoogleNet, MobilenetV2, SqueezeNet, DenseNet201, and Xception) have been utilized to classify segmented breast ultrasound images by transfer learning (TL). A dataset of 375 breast ultrasound images' masks (ground truths), divided as 125 normal, 125 benign, and 125 malignant, has been utilized in training and validation. A dataset of 255 breast ultrasound images' ground truths (masks), divided as 85 normal, 85 benign, and 85 malignant, has been utilized to evaluate the classification accuracy of each CNN model after TL process. Each CNN model's evaluation (over the 255 breast ultrasound images' masks) has represented different accuracy values evaluated. The best accuracy value was for ResNet50 with an accuracy of 97.25 %. The proposed classification scheme for segmented ultrasound images can be regarded as a step that may help in an automatic breast cancer diagnosis system. The produced trained ten CNNs models in this study (input image size: 128 by 128 by 3) are not in their optimum case but it can be considered as a start to any researcher interests in studying the classification of segmented binary images of breast cancers, the ten produced CNNs models are available to researchers at: https://www.kaggle.com/mohammedtgadallah/ten-cnns-classify-segmented-breast-ultrasound.