Tumor Segmentation in Breast Ultrasound Image by Means of Res Path Combined with Dense Connection Neural Network

Abstract

Over the past few years, researchers have demonstrated the possibilities to use the Computer-Aided Diagnosis (CAD) to provide a preliminary diagnosis. Recently, it is also becoming increasingly common for doctors and computer practitioners to collaborate on developing CAD. Since the early diagnosis of breast cancer is the most critical step, a precise segmentation of breast tumor with accurate edge and shape is vital for accurate diagnoses and reduction in the patients’ pain. In view of the deficient accuracy of existing method, we proposed a novel method based on U-Net to improve the tumor segmentation accuracy in breast ultrasound images. First, Res Path was introduced into the U-Net to reduce the difference between the feature maps of the encoder and decoder. Then, a new connection, dense block from the input of the feature maps in the encoding-to-decoding section, was added to reduce the feature information loss and alleviate the vanishing gradient problem. A breast ultrasound database, which contains 538 tumor images, from Xinhua Hospital in Shanghai and marked by two professional doctors was used to train and test models. We, using ten-fold cross-validation method, compared the U-Net, U-Net with Res Path, and the proposed method to verify the improvements. The results demonstrated an overall improvement by the proposed approach when compared with the other in terms of true-positive rate, false-positive rate, Hausdorff distance indices, Jaccard similarity, and Dice coefficients.