UNeSt: A fast segmentation network for colorectal polyps based on MLP and deep separable convolution

Abstract

In medical image segmentation, conventional methodologies based on the UNet method often focus on improving the network performance but overlook the parameters and computational complexity. Due to the limitation of computing resources, these methods can hardly be applied in the landscape of point-of-care (PoC) applications. This study presents UNeSt, a rapid segmentation network tailored for colorectal polyps. The architectural foundation of UNeSt hinges upon the synergistic integration of the depth separable convolutional layer (DSC) and multilayer perceptron (MLP). UNeSt achieves an innovative fusion of these components, resulting in a substantial reduction in model parameters and computational complexity, which is concomitant with a remarkable enhancement in inference speed. Specifically, UNeSt incorporates the convolutional block attention module (CBAM) within the convolutional encoder to extract channel and spatial information proficiently. Furthermore, we introduce an attention mechanism to address the positional information discrepancies introduced in the MLP stage. This comprehensive approach contributes significantly to the augmentation of accuracy in colorectal polyp segmentation. Finally, UNeSt employs skip connections between various levels of encoders and decoders, thereby mitigating information loss problems. In the context of this investigation, UNeSt underwent rigorous evaluation using a demanding polyp segmentation dataset. Relative to UNeXt, a widely employed and exceedingly lightweight network model, the proposed model in this study exhibits a noteworthy reduction with 1.6x fewer parameters, a 2.5x decrease in computational complexity (measured in GFLOPs), and a 1.9x acceleration in inference speed.