Stripe rust, caused by Puccinia striiformis f. sp. tritici, and leaf rust, caused by Puccinia triticina, are among the most damaging wheat diseases worldwide, and their urediniospores serve as the primary inoculum for airborne dispersal. Accurate discrimination of these spores is important for disease forecasting, yet the two spore types are highly similar under light microscopy. For field-monitoring scenarios in which chemical preprocessing is impractical, we developed RustSporeSeg, a deep learning-based semantic segmentation model for direct pixel-level discrimination of stripe rust and leaf rust urediniospores in untreated microscopic images acquired under standardized laboratory conditions. The model integrates attention-based foreground enhancement with multilevel feature refinement and achieved an mIoU of 80.32% and a mean F1-score of 88.45% on the held-out untreated test set. We further evaluated hydrochloric acid (HCl) treatment as a separate laboratory reference condition. Under this condition, several conventional convolutional models achieved mIoU values above 91% and mean F1-scores above 95%, with the best-performing model reaching an mIoU of 95.70% and a mean F1-score of 97.78%. To support further research, a fine-grained urediniospore image dataset encompassing both pathogens, together with the model code, has been made publicly available.These results demonstrate the feasibility of pathogen-specific urediniospore segmentation under the tested acquisition conditions and provide a basis for future validation in automated spore-monitoring workflows.