Dynamically Receptive Field Networks for Image Segmentation

To address the issues of information loss during down-sampling and the insufficient extraction of the multi-scale local semantic information caused by the fixed-size convolutional kernels when applying a U-Net-based method to realize medical image segmentation, a dynamically receptive field network (DRFN) is proposed. Firstly, the network enhances the reducibility of the semantic information during decoding process by introducing the feature progressively cascaded integration (FPCI) module to extract multi-scale local semantic features from the encoder and assigns these features to the decoder. On this basis, the local vision adjustment (LVA) module is designed to strengthen the ability of the fixed-size convolution kernel to obtain contextual semantic information in a large receptive field. The experimental results on the ISIC2018 and BUSI cancer image segmentation datasets show that the IoU score of the proposed method reaches 83.92±0.26 and 70.45±1.70 while the Dice score reaches 91.09±0.23 and 83.39±1.15, which outperforms the state-of-the-art medical image segmentation methods.