Super U-Net: A modularized generalizable architecture

作者：

Highlights：

• A novel convolutional neural network termed ‘Super U-Net” for medical image segmentation.

• A fusion upsampling module that recalibrates feature maps prior to concatenation.

• A dynamic receptive field module that allows the network to determine the correct kernel size for the current segmentation task.

• Comparative experiments were performed on the super U-Net, seven U-Net variants, and two non-U-Net segmentation architectures on the DRIVE, CHASE DB1, Kvasir-SEG, and ISIC 2017 datasets.

摘要

•A novel convolutional neural network termed ‘Super U-Net” for medical image segmentation.•A fusion upsampling module that recalibrates feature maps prior to concatenation.•A dynamic receptive field module that allows the network to determine the correct kernel size for the current segmentation task.•Comparative experiments were performed on the super U-Net, seven U-Net variants, and two non-U-Net segmentation architectures on the DRIVE, CHASE DB1, Kvasir-SEG, and ISIC 2017 datasets.

论文关键词：Image segmentation,U-Net,Dynamic receptive field,Fusion upsampling

论文评审过程：Received 17 August 2021, Revised 10 February 2022, Accepted 24 March 2022, Available online 1 April 2022, Version of Record 7 April 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108669