Sunday 02 March 2025
Medical image segmentation, a crucial task in diagnosing and treating various diseases, has long been a challenge for researchers. The ability to accurately identify and isolate specific features within medical images can greatly aid doctors in making informed decisions about patient care. In recent years, deep learning models have shown great promise in tackling this problem, but their limitations have also become apparent.
Traditional convolutional neural networks (CNNs) excel at capturing local patterns, but struggle with modeling long-range dependencies and global context. Meanwhile, transformer-based models, which have gained popularity for their ability to model long-range relationships, demand significant computational resources and data volumes. To address these limitations, researchers have proposed a novel U-Net architecture called KM-UNet.
KM-UNet combines the strengths of Kolmogorov-Arnold Networks (KANs) and state-space models (SSMs). KANs are designed to efficiently model feature representations, while SSMs enable scalable long-range modeling. By integrating these components, KM-UNet achieves a balance between accuracy and computational efficiency.
The researchers evaluated KM-UNet on five benchmark datasets, including ISIC17, ISIC18, CVC, BUSI, and GLAS. The results show that KM-UNet outperforms state-of-the-art methods in medical image segmentation tasks, with an average IoU of 80.45% and F1 score of 88.63%. Notably, it achieves these impressive results while maintaining a relatively small parameter count (7.35M) and computational cost (17.66 Gflops).
One key aspect of KM-UNet is its Selective-Scan Efficient Multi-scale module, which combines feature extraction and attention mechanisms to effectively model global context and long-range dependencies. This module is instrumental in improving segmentation accuracy, as it enables the model to focus on relevant regions and ignore irrelevant ones.
Another significant contribution of KM-UNet is its ability to enhance model interpretability. By introducing an attention mechanism through the KAN layer, the model can precisely locate key regions and generate activation areas that align better with ground truth masks. This improvement in interpretability is crucial for doctors to understand the reasoning behind the model’s predictions.
The integration of SSMs also allows KM-UNet to efficiently model long-range dependencies, which is particularly important for medical image segmentation tasks where global context plays a critical role.
Cite this article: “KM-UNet: A Novel Architecture for Medical Image Segmentation”, The Science Archive, 2025.
Medical Image Segmentation, Deep Learning, Convolutional Neural Networks, Transformer-Based Models, U-Net Architecture, Kolmogorov-Arnold Networks, State-Space Models, Feature Extraction, Attention Mechanisms, Model Interpretability
Reference: Yibo Zhang, “KM-UNet KAN Mamba UNet for medical image segmentation” (2025).
