Multi-view hybrid encoder U-Net for 3D renal vascular medical image segmentation

Understanding the size, shape, branching angles, and morphological features of blood vessels in human tissue remains challenging. To address this, we propose a multi-view hybrid encoder U-Net for segmenting renal artery vessels. The encoder in this model adopts a hierarchical structure, consisting of an upper encoder and a lower encoder. The upper encoder includes a lightweight convolutional neural network (CNN). In contrast, the lower encoder incorporates either CNN or Transformer modules, forming a flexible hybrid encoding mechanism that enhances feature extraction and representation. The model processes two-dimensional images from three orthogonal views, keeping the original width and height of the input image to preserve spatial resolution. This design helps maintain fine details of high-resolution images, thereby aiding in the capture of subtle features. Experimental results show that the Mean Surface Dice value (MSD) of this method reaches 0.852, and the Dice similarity coefficient (DSC) reaches 0.939 on kidneys not used in training. These findings suggest that the proposed approach is effective for detailed renal artery segmentation, although further validation on diverse datasets and modalities is required to establish its generalizability fully.