Atrous Convolutional Neural Network (ACNN) for Semantic Image Segmentation with full-scale Feature Maps

Deep Convolutional Neural Networks (DCNNs) are used extensively in biomedical image segmentation. However, current DCNNs usually use down sampling layers for increasing the receptive field and gaining abstract semantic information. These down sampling layers decrease the spatial dimension of feature maps, which can be detrimental to semantic image segmentation. Atrous convolution is an alternative for the down sampling layer. It increases the receptive field whilst maintains the spatial dimension of feature maps. In this paper, a method for effective atrous rate setting is proposed to achieve the largest and fully-covered receptive field with a minimum number of atrous convolutional layers. Furthermore, different atrous blocks, shortcut connections and normalization methods are explored to select the optimal network structure setting. These lead to a new and full-scale DCNN - Atrous Convolutional Neural Network (ACNN), which incorporates cascaded atrous II-blocks, residual learning and Fine Group Normalization (FGN). Application results of the proposed ACNN to Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) image segmentation demonstrate that the proposed ACNN can achieve comparable segmentation Dice Similarity Coefficients (DSCs) to U-Net, optimized U-Net and hybrid network, but with significantly reduced trainable parameters due to the use of full-scale feature maps and therefore computationally is much more efficient for both the training and inference.