Under practical Panax notoginseng cultivation scenarios, due to the visual similarities between diseases such as gray mold and plague, as well as small individual targets with complex and variable shapes of diseases such as anthracnose under natural conditions, current methods face a difficult problem of P.notoginseng leaf disease segmentation. A modified Transformer-DCNv2-BlendMask model for P. notoginseng leaf multicategory diseases image segmentation was proposed. To deal with the visual similarity problem and variable shape targets appeared on P. notoginseng leaf disease, a Transformer encoder to capture long-distance dependencies for multiple disease categories was introduced. And the deformable convolution networks v2 (DCNv2) showed a better adaptability of convolutional networks by enabling free-form deformation of the convolution to segment disease with various shape. The model and other instance segmentation models such as BoxInst, ConInst, SOLOv2, Mask R-CNN and YOLO v8-seg on the P. notoginseng leaf disease dataset, which contains multi-category diseases were compared. The results demonstrated the competitive performance of our model, achieving a average precision (AP) of 86.14%, outperforming the baseline BlendMask model by 3.17 percentage points and the previously best-performing Mask R-CNN by 4.37 percentage points. It also exceeded the baseline by 0.16 percentage points, 4.32 percentage points and 4.46 percentage points for the gray mold, plague and anthracnose categories, respectively. Thus, our method provides a robust solution for segmenting shape-variable and visually similar diseases in complex environments, helping to achieve accurate quantification of diseases.