Breast Ultrasound Image Classification and Segmentation Using Convolutional Neural Networks

Due to the shortage and uneven distribution of medical resources all over the world, breast cancer diagnosis and treatment is a fundamental but vital problem, especially in developing countries. Breast ultrasound image classification and segmentation method by using Convolutional Neural Networks (CNN) can be a new efficient solution in early analysis and diagnosis. What’s more, the diagnosing of diversity of cancers is challenge in itself and the training of data-driven based CNN model also highly relay on dataset. In this paper, we first build a breast ultrasound dataset (with 1418 normal and 1182 cancerous samples) labeled by three radiologists from XiangYa Hospital of Hunan Province. And then, we propose a two-stage Computer-Aided Diagnosis (CAD) system to diagnose the breast cancer automatically. Firstly, the system utilize a pre-trained ResNet generated with transfer learning approach to excluded normal candidates, and then use an improved Mask R-CNN model for the accurate tumor segmentation. Experimental results show that the proposed system can achieve 98.72% precision and 98.05% recall for classification, and 85% (1.2% improvement) mAP and 82.7% (3.1% improvement) F1-Measure than the original Mask R-CNN model.