Experimental study demonstrates improved boundary delineation in dermoscopic and endoscopic images, suggesting frequency decomposition optimizes state-space segmentation.
Key Points
To develop and validate MGA-UNet, a frequency-aware multi-scale framework designed to overcome blurry boundaries and capture long-range dependencies in medical image segmentation.
Integrated a Wavelet-Mamba backbone (WMB) to separate low- and high-frequency features, a Gated Multi-scale Aggregation Module (GMAM), and an Adaptive Sparse Attention Module (ASAM) for bottleneck refinement.
Evaluated the framework across three independent runs (random seeds 42, 123, and 2026) on the ISIC2018, ISIC2017, and Kvasir-SEG benchmark datasets.
MGA-UNet achieved mean Dice Similarity Coefficients of 88.92±0.04% on ISIC2018 and 88.01±0.07% on ISIC2017 dermoscopy benchmarks.
MGA-UNet achieved a mean Dice Similarity Coefficient of 85.91±0.04% on the Kvasir-SEG endoscopy dataset, outperforming baseline models including H-VMUNet.