Uncertainty-aware hybrid Mamba network for robust medical image segmentation across modalities

Medical image segmentation requires models that are simultaneously context-aware, computationally efficient, and robust to annotation ambiguity. Conventional CNNs offer strong inductive bias but struggle to propagate information over long spatial ranges, while Transformers enhance global reasoning at the cost of heavy computation and weaker locality priors. To balance these trade-offs, we present an uncertainty-aware hybrid Mamba network for robust medical image segmentation across modalities. Our proposed network is organized into three stages: (i) an encoder built with Mamba– Spectral Convolution (MSC) blocks that jointly capture frequency-domain features and long sequential features via convolutional processing, fast Fourier transform (FFT), and multi-scale Mamba; (ii) a Kolmogorov–Arnold Mamba (KAM) bottleneck that enriches nonlinear representations through group-wise KAN-based transformations alongside state-space sequence modeling; and (iii) a decoder equipped with Uncertainty-aware Mamba Transformer (UMT) blocks that apply uncertainty-guided masking to suppress unreliable boundary regions and emphasize clinically informative patterns using masked efficient self-attention and masked multi-scale Mamba. Extensive evaluations across diverse 2D and 3D datasets (CVC-ClinicDB, BUSI, UDIAT, HIT-FISS, photoacoustic computed tomography (PACT), and AMOS22-MRI) show that our proposed network consistently improves delineation quality, producing cleaner boundaries and reducing false positives in challenging or ambiguous areas compared with other segmentation models.