MF2-Net: A multipath feature fusion network for medical image segmentation

In this paper, we propose a multipath feature fusion convolutional neural network (MF2-Net) with novel and efficient spatial group convolution (SGC) modules with a multipath feature fusion network for the automated segmentation of medical images. The proposed MF2-Net was designed with multiple encoder paths to extract layer-specific multiscale information. Each encoder path employs SGC modules composed of stacked asymmetric kernels of different sizes (1 and 1). In the SGC modules, the context details of high-level features are encoded at varying scales, and neighbor feature information is incorporated with higher precision. In addition, the encoded features fused at the bottleneck layer capture abundant semantic features from input images. Furthermore, the guided block mechanism is used to refine the segmentation boundaries at the decoder stage by integrating the skip connection from the encoder stage. We verified that the SGC modules in a multipath feature fusion network improve the segmentation accuracy with fewer learnable parameters. Experimental results demonstrated that the proposed model outperformed existing medical image segmentation methods by an average score of 0.97 on publicly available datasets.