Hippocampus segmentation in MRI using side U-net model

Abstract

Convolutional neural networks (CNN) have been applied in medical image analysis over the past few years. U-Net architecture is one of the most well-known CNN architectures in many different medical image segmentation tasks. However, it is hard to capture subtle local features because of its limitations in standard convolution layers and one output prediction. In addition, some objects like hippocampus in the biomedical image occupies an only small area which increases the difficulty of segmentation. In this manuscript, we present an architecture, called Side U-Net, which addresses these challenging problems. In the condition of giving unbalanced class images, Side U-Net outperforms the U-Net by upgrading loss function and capturing more important local features using multiple side outputs. And the experimental results verified our method and demonstrated that our method outperformed the U-Net model over 0.75% in terms of dice score and in the same threshold of classification, our model has a higher TPR (True Positive Rate) when evaluated in ADNI dataset.