Improved U-Net Model for Nerve Segmentation

Noticeable gains in computer vision have been made as a result of the large-scale datasets and deep convolutional neural networks (CNNs). CNNs have been used in a wide variety of tasks, for instance, recognition, detection, and segmentation. Recently, due to the open medical images datasets, CNNs have been used in Computer Aided Detection (CADe) to help doctors diagnose lesion. In this work, we present an end-to-end method based on CNNs for automatical segmentation from medical images. The proposed network architecture is similar to U-Net, which consists of a contracting path and an expansive path. However, we take advantage of inception modules and batch normalization instead of ordinary convolutional layers, which reduce the quantity of parameters and accelerate training without loss of accuracy. In addition, we confirm Dice coefficient as loss function rather than binary cross entropy. We use this model to segment nerve from ultrasound images and achieve a better performance.