Analysis of Different Encoder-decoder-based Approaches for Biomedical Imaging Segmentation

Recently, CNNs (convolutional neural networks) have been widely used in the field of medical image segmentation. In particular, the encoder-decoder architectures represented by U-Net have achieved state-of-art segmentation effects and inspired many more elaborated networks, which adopt newer and more advanced network designs. To our knowledge, the comprehensive and detailed comparison among these improved versions from a multiplicity of points of view has not been conducted up to now. With U-Net as the baseline, we select the other four typical improvements for U-Net. For higher reliability, we finish the task of segmentation on four datasets and more experiments are performed to test the performance in various conditions. Finally, we evaluate their performance using multiple evaluation metrics. We find that attention U-Net achieves the best segmentation results in terms of F1-score but also owns the most trainable parameters and is most time-consuming. As training images decrease, the original U-Net is most robust even only less than 5 training samples are available. Besides, for any networks, adding auxiliary loss function with small weighting such as 0.01 or 0.01 whatever the cross-entropy loss and the dice-coefficient loss for the other one is beneficial as well.