Development of a Deep-Learning-Based Method for Breast Ultrasound Image Segmentation

Breast cancer is one of the deadliest cancers that cause women death globally. Ultrasound imaging is one of the commonly used diagnostic tools for detection and classification of breast abnormalities. In the past decades, computer-aided diagnosis (CAD) systems have been developed to improve the accuracy of diagnosis made by radiologists. In particular, automatic breast ultrasound (BUS) image segmentation is a critical step for cancer diagnosis using CAD. However, accurate tumor segmentation is still a challenge as a result of various ultrasound artifacts. This paper developed a novel segmentation framework based on deep learning architecture u-net, for breast ultrasound imaging. U-net is a convolutional neural network architecture designed for biology image segmentation with limited training data. It was originally proposed for neuronal structure segmentation in microscopy images. In our work, we modified and improved the method for BUS image segmentation. On a database of 221 BUS images, we first applied pre-processing techniques including contrast enhancement and speckle reduction to improve the image quality. Then the u-net model was trained and tested through two-fold cross-validation. In order to increase the size of training set, data augmentation strategies including rotation and elastic deformation were applied. Finally, a post-processing step that removed noisy region(s) from the segmentation result finalized the whole method. The area error metrics, dice coefficient, and similarity rate were calculated to evaluate the performance on the testing sets. We compared our method with another two fully automatic segmentation methods on the same dataset. Our method outperformed the other two significantly with the dice coefficient = 0.825 and similarity rate = 0.698. Experiment results showed that the modified u-net method is more robust and accurate in breast tumor segmentation for ultrasound images.