Purpose: Deep learning has recently shown its outstanding performance in biomedical image semantic segmentation. Most biomedical semantic segmentation frameworks comprise the encoder-decoder architecture directly fusing features of the encoder and the decoder by the way of skip connections. However, the simple fusion operation may neglect the semantic gaps which lie between these features in the decoder and the encoder, hindering the effectiveness of the network.
Methods: Dense gate network (DG-Net) is proposed for biomedical image segmentation. In this model, the Gate Aggregate structure is utilized to reduce the semantic gaps between features in the encoder and the corresponding features in the decoder, and the gate unit is used to reduce the categorical ambiguity as well as to guide the low-level high-resolution features to recover semantic information. Through this method, the features could reach a similar semantic level before fusion, which is helpful for reducing semantic gaps, thereby producing accurate results.
Results: Four medical semantic segmentation experiments, based on CT and microscopy images datasets, were performed to evaluate our model. In the cross-validation experiments, the proposed method achieves IOU scores of 97.953%, 89.569%, 81.870% and 76.486% on these four datasets. Compared with U-Net and MultiResUNet methods, DG-Net yields a higher average score on IOU and Acc.
Conclusion: The DG-Net is competitive with the baseline methods. The experiment results indicate that Gate Aggregate structure and gate unit could improve the performance of the network by aggregating features from different layers and reducing the semantic gaps of features in the encoder and the decoder. This has potential in biomedical image segmentation.
Keywords: Biomedical images; Convolution neural networks; Deep learning; Semantic segmentation.