Segmentation in Skeletal Scintigraphy Images using Convolutional Neural Networks

In this work we have addressed the task of segmentation in skeletal scintigraphy images with deep learning models, where we research different approaches to convert convolutional neural networks designed for classification tasks to powerful pixel wise predictors. We explore different network architectures where two primary research paths have been followed. Firstly, an encoder-decoder architecture which aims to extract dense features in the first part of the network -- which works as a feature encoder -- and then up-sample these dense features to restore the original image resolution and perform pixel-wise predictions. This technique has shown great promise in other experiments of segmentation in medical images. This general architecture has been pitted against an entirely different approach, which works with expansions of the convolutional kernels, rather than sub-sampling through pooling layers, known as convolutions ``atrous'' or dilated convolutional kernels. While the atrous approach has been explored in different studies for the problem of semantic image segmentation for outdoor and indoor scenes with a large amount of classes it has yet to be tried in the medical imaging field. When compared to the encoder-decoder architecture we see that the convolutional neural networks atrous outperform the former in almost every way. We observe that the most promising atrous model generated a test error of 0.0659 on segmentations on the left scapula, which is a reduction of 50.67\% on the test error as compared to the most powerful encoder-decoder model. The atrous model also managed to reduce the amount of parameters by a factor 100 and more than halve the required training time per epoch