Medical image segmentation (MIS) is crucial for disease diagnosis, but remains challenging due to the variations in lesion morphology and complex anatomical structures. In MIS tasks, the Mamba architecture is highly regarded for its capability to capture global information efficiently, with linear computational complexity, while ensuring high segmentation accuracy. Building upon this foundation, we propose MSFM-UNet, a novel U-shaped hybrid network that combines the strengths of CNNs and the Mamba model, utilizing a Cross-Conv Block (CCB), a Random Multi Scan (RMS), and an Adaptive Frequency Domain Filtering Module (AFDFM) to enhance feature extraction, spatial representation, and frequency domain understanding. Specifically, CCB bridges CNN and Mamba feature extraction by leveraging vertically and horizontally learnable convolutional kernels to align diverse feature representations, RMS integrates vertical, horizontal, and diagonal slicing with randomly selected scanning modes, capturing more comprehensive spatial information and improving model robustness. Additionally, in the bottleneck layer, AFDFM is proposed to extract frequency domain features essential for modeling tissue and organ structures. Extensive experiments on the CVC-ClinicDB, ISIC2017, and ISIC2018 datasets demonstrate that MSFM-UNet achieves superior segmentation accuracy compared to existing methods without compromising computational efficiency.