Segmentation of Medical Ultrasound Images Using Convolutional Neural Networks with Noisy Activating Functions

The attempts to segment medical ultrasound images have had limited success than the attempts to segment images from other medical imaging modalities. In this project, we attempt to segment medical ultrasound images using convolutional neural networks (CNNs) with a group of noisy activation functions which have recently been demonstrated to improve the performance of neural networks. We report on the segmentation results using a U-Net-like CNN with noisy rectified linear unit (NReLU) functions, noisy hard sigmoid (NHSigmoid) functions, and noisy hard tanh (NHTanh) function on a small data set.