Deep learning in general, and segmentation problems in particular, are the subject of much research done in the field of learning systems. The rising power of computers enables the implementation of complex networks containing thousands of variables in order to create optimal architecture for such problems. However, there are several approaches to dealing with segmentation problems, each of which has its own advantages and disadvantages.
This project deals with a segmentation problem in which we are required to locate a particular element – the Brachial Plexus (a network of nerves in the human body), from a given pool of ultrasound images.
Most of the project’s efforts have been to find a suitable architecture that will best locate the BP. The architectures tested are based on convolution layers, which have many advantages in image processing tasks. In addition to the convolution layers, additional layer types are used, such as activation functions that enable the network to learn more complex functions, pooling layers to reduce dimension, bilinear interpolation functions to enable dimension enhancement, and more.
Another element examined in the project is the effect of combining a segmentation network and a binary classifier that determines whether a particular image contains the BP. The classifier is required to perform a task that is less complex than the segmentation network, because it needs to determine a binary answer without providing details such as the size and position of the BP in the image. In general, the segmentation network alone can also determine the answer to the question of the BP presence in the image. However, it turns out that splitting the tasks into two different components makes it possible for each network to perform their task optimally – indeed, the combination of a binary classifier improved the system’s overall performance.
Pre-processing of the image also has a major impact on the success of the network. It turns out that actions such as reducing the dimensions of the image and filtering it simplifies the task in a way that allows the network to learn better. During the project, these tools were widely used as well as other types of layers, such as the dilated convolution layer.