Light-VM2D-UNet: A Lightweight UNet Enhanced with Mamba2D for Medical Image Segmentation

This paper addresses the fundamental trade-off between global long-range dependency modeling and local detail preservation in lightweight medical image segmentation models. Through systematic analysis of the limitations in spatial dependency capturing exhibited by the parallel vision Mamba (PVM) component in UltraLight VM-UNet procedure, we propose a novel PVM2DRes Block. This module integrates a 2D wavefront scanning strategy into the PVM unit while incorporating local contextual enhancement during parallel state space updates. The proposed approach enables multi-directional and multi-scale global-local feature interactions with maintained linear computational complexity. The resulting Light-VM2D-UNet achieves superior performance with merely 0.12 M parameters and 0.07 GFLOPs, demonstrating consistent improvements in accuracy and mIoU metrics across three benchmark datasets: ISIC-2017, ISIC-2018, and CVC-ClinicDB. Extensive experiments validate that our method not only enhances the modeling capability for long-range feature dependencies and fine-grained local details but also establishes the generalizability of integrating state space models with attention mechanisms. This work provides a more efficient and accurate technical pathway for intelligent diagnostic systems in resource-constrained environments.