A high-performance nuisance SMS recognition model

SCANeXt: Enhancing 3D medical image segmentation with dual attention network and depth-wise convolution

“…The SCANeXt utilizes a hierarchical design that reduces feature resolution by a factor of two at each stage, efficiently capturing both local and global context information. Our SCANeXt framework distinguishes itself from the recently proposed TransUNet [60] and TFCNs [64] . TransUNet primarily focuses on integrating the U-Net architecture with Transformers to capture local features and global context information.…”

“…This enables the interleaving of convolutional and transformer-based blocks in the encoder and decoder, fully leveraging their respective advantages in feature extraction. TFCNs [64] introduce the Convolutional Linear Attention Block (CLAB), which encompasses two types of attention: spatial attention over the image's spatial extent and channel attention over CNN-style feature channels. The aforementioned methods simply rearrange the CNN and transformer modules within the encoder and decoder structures of UNet, without introducing any novel modules, resulting in limited improvements in segmentation performance.…”