Fully convolutional neural network for semantic segmentation on CT scans of pigs

Identifying the shape and location of structures within medical images is useful for purposes such as diagnosis and research. This is a cumbersome task if done manually. Recent advances in computer vision and in particular deep learning have made it possible to automate this task to such an extent that it is comparable to human level performance. This thesis reviews the components used to construct a fully convolutional neural network for semantic segmentation. It then proposes a modified network architecture based on an existing state-of-the-art fully convolutional neural network called U-net. The architecture is applied to a binary classification problem involving computed tomography scans of pigs provided by Norsvin SA. The goal is to classify each pixel in the scans as either "a part of the pig which is edible" or "background" which means everything that is not in the edible class. Each computed tomography scan is too large for the network to process at once. Part of the thesis is therefore devoted to investigating approaches for feeding the information in the scans to the proposed network. The network is trained on 238 scans and evaluated on 37 scans. The evaluation is done quantitatively using the index over union metric and qualitatively through manual inspection of segmented images. The results show that the best performing network on average obtains an index over union score of 0.962 when given a scan for segmentation.