MS-UNet: A Hybrid Network with a Multi-Scale Vision Transformer and Attention Learning Confusion Regions for Soybean Rust Fungus

Soybean rust, caused by the fungus Phakopsora pachyrhizi, is recognized as the most devastating disease affecting soybean crops worldwide. In practical applications, performing accurate Phakopsora pachyrhizi segmentation (PPS) is essential for elucidating the morphodynamics of soybean rust, thereby facilitating effective prevention strategies and advancing research on related soybean diseases. Despite its importance, studies focusing on PPS-related datasets and the automatic segmentation of Phakopsora pachyrhizi remain limited. To address this gap, we propose an efficient semantic segmentation model named MS-UNet (Multi-Scale Confusion UNet Network). In the hierarchical Vision Transformer (ViT) module, the feature maps are down-sampled to reduce the lengths of the keys (K) and values (V), thereby minimizing the computational complexity. This design not only lowers the resource demands of the transformer but also enables the network to effectively capture multi-scale and high-resolution features. Additionally, depthwise separable convolutions are employed to compensate for positional information, which alleviates the difficulty the ViT faces in learning robust positional encodings, especially for small datasets. Furthermore, MS-UNet dynamically generates labels for both hard-to-segment and easy-to-segment regions, compelling the network to concentrate on more challenging locations and improving its overall segmentation capability. Compared to the existing state-of-the-art methods, our approach achieves a superior performance in PPS tasks.