Describing ultrasound video content using deep convolutional neural networks

We address the task of object recognition in obstetric ultrasound videos using deep Convolutional Neural Networks (CNNs). A transfer learning based design is presented to study the transferability of features learnt from natural images to ultrasound image object recognition which on the surface is a very different problem. Our results demonstrate that CNNs initialised with large-scale pre-trained networks outperform those directly learnt from small-scale ultrasound data (91.5% versus 87.9%), in terms of object identification.