A Modified nnWNet Architecture With Multi-Scale and State-Space Enhancements for Medical Image Segmentation

We propose WNet2D, a lightweight enhancement to nnWNet for medical image segmentation that aims to preserve fine boundaries while enforcing global consistency under small-cohort settings. WNet2D introduces three complementary modules: MS-LSB for shallow-edge modeling, E-GSB for mid-level semantic aggregation, and a single bottleneck M-GSB (Mamba) for deep long-range dependencies. On Kvasir-SEG and PH2, WNet2D achieves absolute Dice/Jaccard gains of +1.22/+0.58 and +0.27/+0.20 percentage points over nnWNet, respectively; on ISIC 2017 the gains are small (Dice +0.07, Jaccard +0.17), and on DRIVE performance is broadly comparable (Dice −0.72, Jaccard +0.14). Despite the higher accuracy, the network remains compact (4.00M parameters) and, under a unified FP32, 512×512, batch-1 protocol across multiple GPUs, delivers ≈42–46% lower latency and ≈85% lower peak GPU memory than nnWNet. The design choice—attention-free shallow/mid stages paired with one deep state-space block—improves boundary fidelity and long-range consistency without added deployment complexity. Overall, WNet2D offers a favorable accuracy–efficiency trade-off and practical deployability across modalities and hardware.