An efficient deep neural network based abnormality detection and multi-class breast tumor classification

Rakesh Chandra Joshi, Divyanshu Singh, Vaibhav Tiwari, Malay Kishore Dutta

Abstract: Breast tumor is one of the major cause of death among women all over the world. Ultrasound imaging-based breast abnormality detection and classification play a vital role to develop an automatic computer-aided diagnostic system. In this paper, deep learning technology is integrated with ultrasound images for pre-screening of breast cancer. Two breast ultrasound image datasets are trained on different deep-learning architectures with image augmentation. Convolutional neural network extracts the features from training ultrasound images which are fine-tuned for multiple iterations. The experimental outcomes indicate accurate and rapid prediction performance on the test dataset of 2D B-mode ultrasound images, signifying a promising approach for assistance to radiologists in clinical applications with the use of deep learning. Results demonstrate the proposed method attains an accuracy, sensitivity, and specificity of 96.31%, 92.63%, and 96.71% respectively. About 12 B-mode 2D ultrasound image frames can be processed per second which marks it as a highly efficient system. The proposed method gives better performance compared to other methods which shows its effectiveness in real-time computer-aided diagnosis of breast tumor and benign-malignant classification.