Medical image segmentation model based on double decoder
Since the target area scales of medical images were different,and samples of labeled medical images were few,a dual decoder medical image segmentation model DDS-UNet was proposed.To be more specific,the DDS-UNet model used Swin Transformer module to construct the encoder,to extract multi-scale features of medical images.The decoder 1 took advantage of Swin Transformer module for global and remote semantic feature extraction to recover and aggregate the corresponding scale feature information of the encoder output step by step during the upsampling process.The decoder 2 made use of the local feature extraction advantage of convolutional neural networks(CNN)to recover the spatial information of medical images step by step during the upsampling process.The feature fusion module used the cavity convolution to decompose the deep semantic feature information output by the encoder,and collaboratively fused the multi-scale feature information output by the double decoders in the upsampling process,so as to reconstruct the spatial details of the target region of the medical image.The experimental results of spine and brain glioma image segmentation showed that the DDS-UNet model had significant abilities on feature extraction and segmentation for the target region.The ablation experiment further verified the effectiveness of the DDS-UNet model for medical image segmentation.
medical image segmentationdouble decoderSwin Transformeratrous convolutionmulti-scale feature fusion
NETL
NSTL
万方数据
