Computational Intelligence Artificial Intelligence
ComFormer : A Hierarchical Multi‐Scale Hybrid Transformer for Precise Medical Image Segmentation
Abstract
ABSTRACT Accurate medical image segmentation requires both global context modeling and precise boundary recovery. Existing CNN‐only and Transformer‐only solutions often struggle to balance these requirements under limited training data and practical efficiency constraints. We propose ComFormer, where “Com” denotes complementary context modeling : global dependencies and local spatial details are jointly modeled by combining Efficient Dual Attention, External Attention, Graph Convolution Block, and a Multi‐Scale Feature Bridge (MSFB). We further provide internal ablations, complexity profiling, and boundary‐focused analyses. Across Synapse, ISIC‐2018, and Kvasir‐SEG, ComFormer consistently delivers competitive or superior segmentation quality. On Synapse, ComFormer achieves 83.46% DSC and 16.89 mm HD95. Additional analyses show a measurable trade‐off between region‐overlap and boundary‐region metrics across MSFB variants while maintaining practical computational cost (26.05M parameters and 6.15G FLOPs under input). ComFormer provides a practical hybrid segmentation framework with explicit evidence on generalization, efficiency, and spatial‐detail preservation, making it suitable for clinically relevant image segmentation scenarios.