FFUNet: A novel feature fusion makes strong decoder for medical image segmentation

Convolutional neural networks (CNNs) have strong ability to extract local features, but it is slightly lacking in extracting global contexts. In contrast, transformers are good at long-distance modelling due the self-attention mechanisms while its performance localization limited. On other hand, feature gap between an encoder and decoder also challenging for a U-shaped network, which adopts plain skip connection. Inherited from convolutional transformers, FFUNet, hybrid network structure with novel module named Feature Fusion Module (FFM) proposed medical image segmentation. The FFM consisting of Attention Selection, Cross Offset Generation Deformable Convolution Layer, aims replace original connection alleviate ambiguous semantic information more powerful segmentation network. Extensive experiments demonstrate that FFUNet has amazing gains on Synapse dataset. addition, consistent improvements achieved across four popular datasets CNN-based or transformer-based networks, illustrate method advantages generalization compactness.