Tongue Image Segmentation Method Based on the VDAU-Net Model

Accurate segmentation of the tongue is fundamental to tongue diagnosis, yet existing methods are predicated on images captured under standard conditions. In practical tongue diagnosis, such controlled environments are not always feasible. This study, therefore, aims to develop a segmentation method capable of effectively isolating the tongue region from complex backgrounds, thereby enhancing the precision and efficiency of tongue diagnosis. Initially, we constructed a novel tongue image segmentation dataset with complex backgrounds, which comprises 1,600 images obtained through data augmentation techniques. Subsequently, We introduce VDAU-Net, an enhanced U-Net-based model for tongue segmentation, featuring three core innovations: 1) use of VGG16 as the encoder (V) to extract multi-level semantic features; 2) implementation of a dual-attention mechanism (DA) that integrates Position Convolution Block Attention Module (PCBAM) – focusing on tongue fissures and coating textures-with Shifted Window Attention (SWA) to enhance the continuity of tongue contours; and 3) realization of cross-scale feature fusion within the U-shaped encoder-decoder framework. Together, these innovations enable pixel-precise tongue segmentation against complex backgrounds. The VDAU-Net demonstrated exceptional performance on our dataset, achieving a Mean Intersection over Union of 96.73%. It effectively captured detailed tongue features and accurately localized the tongue region, even under challenging conditions. The VDAU-Net, by integrating advanced attention mechanisms and transfer learning, can effectively handle complex backgrounds and variable imaging conditions, making it a powerful tool for intelligent tongue diagnosis systems. Future research will focus on expanding the dataset and refining the algorithm to enhance its applicability across diverse environments.