Segmentation tools help doctors identify and delineate tumors, lesions, and other abnormalities in medical images. However, noise, redundant information, and blurred boundaries can make automated segmentation difficult. Researchers from Tongji University in China and the University of Toyama in Japan developed V-UNet, a new model designed to capture global context and fine local details. Tests on four public datasets demonstrated competitive segmentation performance with relatively low computational requirements.
When doctors examine medical images, obtained from X-rays, MRI scans, CT scans, or ultrasound, they often need to identify and accurately trace the boundaries of tumors, lesions, or other types of abnormal tissue. Known as image segmentation, this step is essential for accurate diagnosis and treatment. Since performing manual segmentation across large volumes of images is time-consuming and vulnerable to fatigue-related errors, machine learning tools are increasingly being used to assist with this task.
However, medical images still present a difficult challenge for automated segmentation. Lesions and tumors can vary greatly in size and appearance, often showing blurred boundaries. To work properly, a segmentation model needs to recognize fine local details to identify boundaries while also capturing long-range relationships between different parts of an image. Convolutional neural networks (CNNs) are effective at extracting local features but struggle with these long-range relationships. Transformers, on the other hand, can capture them more effectively, but their attention mechanisms can become computationally expensive as image size increases. Newer models based on the 'Mamba' architecture offer a promising alternative, but they remain vulnerable to noise and redundant information that is often present in medical images.
To address this gap, a research team led by Professor Jiujun Cheng from Tongji University, China, and Professor Shangce Gao from the University of Toyama, Japan, has developed V-UNet, a novel Mamba-UNet architecture for medical image segmentation, which combines complementary feature-extraction approaches designed to address these limitations. Their paper was published in Volume 13, Issue 6 of IEEE/CAA Journal of Automatica Sinica on July 3, 2026.
The core of their method is a new building block called "high-low-order feature fusion visual state-space (HL-VSS)," which works in two stages. First, it uses multi-scale spatial convolution to extract useful features at different image scales while reducing the influence of noise. Then, a feature fusion step combines two complementary paths: one that captures broad spatial relationships across the image, and another that focuses on fine local texture and detail. The two paths are then combined, so the model benefits from both perspectives at once. This entire module is built into a custom U-shaped network architecture, a widely used design in medical image segmentation.
"Feature fusion not only preserves critical structural features but also effectively suppresses background noise, thereby further enhancing the distinction between the target regions and the true boundary information," explains Prof. Cheng, "This method maximally reduces the impact of noise and redundant information, thereby substantially improving segmentation performance and enhancing the potential for clinical applications in medical image segmentation."
The team tested V-UNet, against several existing state-of-the-art models across four public medical imaging datasets, covering skin lesion images and breast ultrasound images. V-UNet achieved competitive segmentation performance compared with existing models while maintaining relatively low numbers of parameters and FLOPs. This is an especially important consideration for hospitals and clinics with limited computing power. "The comprehensive series of experiments we conducted and our comparative results ultimately confirm that V-UNet is a superior model," notes Prof. Gao.
Taken together, the findings point to a possible route toward more efficient and robust AI-assisted medical image analysis. Better segmentation could help computer-aided systems more reliably delineate abnormal regions for subsequent analysis, although further research is needed to determine the model's performance in real-world clinical settings. Going forward, the researchers plan to explore how frequency noise, a type of interference not addressed in this study, affects segmentation accuracy, aiming to further refine the model's reliability for real-world applications.
