RMIS-Net: a fast medical image segmentation network based on multilayer perceptron

Medical image segmentation, a pivotal component in diagnostic workflows and therapeutic decision-making, plays a critical role in clinical applications ranging from pathological diagnosis to surgical navigation and treatment evaluation. To address the persistent challenges of computational complexity and efficiency limitations in existing methods, we propose RMIS-Net-an innovative lightweight segmentation network with three core components: a convolutional layer for preliminary feature extraction, a shift-based fully connected layer for parameter-efficient spatial modeling, and a tokenized multilayer perceptron for global context capture. This architecture achieves significant parameter reduction while enhancing local feature representation through optimized shift operations. The network incorporates layer normalization and dropout regularization to ensure training stability, complemented by Gaussian error linear unit (GELU) activation functions for improved non-linear modeling. To further refine segmentation precision, we integrate residual connections for gradient flow optimization, a Dice loss function for class imbalance mitigation, and bilinear interpolation for accurate mask reconstruction. Comprehensive evaluations on two benchmark datasets (2018 Data Science Bowl for cellular structure segmentation and ISIC-2018 for lesion boundary delineation) demonstrate RMIS-Net's superior performance, achieving state-of-the-art metrics including an average F1-score of 0.91 and mean intersection-over-union of 0.82. Remarkably, the proposed architecture requires only 0.03 s per image inference while achieving 27× parameter compression, 10× acceleration in inference speed, and 53× reduction in computational complexity compared to conventional approaches, establishing new benchmarks for efficient yet accurate medical image analysis.