Mamba-SUDS: Integrating Selective State Space Models into a Simplified U-Net Architecture with Depthwise Separable Convolutions for Medical Image Segmentation

Medical image segmentation is essential for clinical diagnosis and treatment planning. Despite the notable performance of convolutional neural networks (CNNs), they encounter difficulties in capturing long-range relationships crucial for precise boundary delineation. Recent state space models, especially Mamba, provide an efficient substitute for attention mechanisms with linear computational complexity. This paper presents Mamba-SUDS, which incorporates selective state space models into the lightweight SUDS (Simplified U-Net Architecture with Depthwise Separable Convolutions) framework for 2D and 3D medical image segmentation. Our approach integrates bidirectional Mamba layers at the encoder–decoder bridge and at skip connections to capture global context while preserving computational efficiency. We evaluate Mamba-SUDS on three datasets: ISIC 2018 (skin lesions), CVC-ClinicDB (polyps) and BraTS 2023 (brain tumors), and compare it against a matching baseline in which the Mamba modules are removed and all other factors held fixed. On the official ISIC 2018 test set Mamba-SUDS reaches a mean IoU of 0.808 ± 0.002 over three seeds with 11.77 M parameters, against 0.787 ± 0.003 for the matching baseline; on CVC-ClinicDB the ordering reverses, so the benefit is not uniform across datasets. On BraTS 2023, using full-volume sliding-window inference, it attains Dice scores of 0.889, 0.826 and 0.809 for whole tumor, tumor core and enhancing tumor. We report efficiency in terms of parameters, multiply–accumulate operations, latency and peak memory, and document the evaluation protocol for every dataset in full.