Segmentation of Breast Ultrasound Images using Densely Connected Deep Convolutional Neural Network and Attention Gates

Abstract

Ultrasound imagining modality is a popular complementary technique for diagnosing breast cancer. A standardized reporting process called Breast imaging reporting and data system (BI-RADS) is used to categorize breast cancer. The BI-RADS scale uses several features of lesions based on the ultrasound images, which makes the quality of the diagnosis highly dependent on the experience of the radiologist. Radiologists use Computer-Aided Diagnosis (CAD) system to help in the detection of lesions. The accuracy of a CAD system depends greatly on the segmentation stage of the system. To increase the reliability of the diagnosis, we propose a solution based on a densely connected deep convolutional neural network and attention gates, called Attention U-DenseNet. Attention U-DenseNet is an architecture to do semantic segmentation of the lesions from Breast Ultrasound (BUS) images based on the U-Net, DenseNet, and attention gates. Convolutional layers of the U-Net are made densely connected using dense blocks to help to learn complex patterns of the BUS image which is usually noisy and contaminated with speckles. This architecture (U-DenseNet) produced an F-score of 0.63 compared to the U-Net model with an F-score of 0.49. Furthermore, to localize the segmentation by learning salient features, attention gates are added to the U-DenseNet architecture (Attention U-DenseNet). Attention U-DenseNet performed even better compared to U-DenseNet, by improving the F-score to 0.75. Finally, a per-image regularised binary cross-entropy is employed to penalize false negatives more than false positives, since the region of interest is small.