SEM PathoDiff: Latent Diffusion-Augmented Scanning Electron Microscope Images of Pathogenic Bacteria Vibrio cholerae, Streptococcus pneumoniae, and Staphylococcus aureus

Dataset is designed for automated pathogen detection using deep learning and machine learning, providing a reproducible benchmark for bacterial image classification under low-data conditions. This dataset addresses the scarcity of publicly available high-quality SEM images of clinically significant bacteria and enables scientifically defendable evaluation of AI models for pathogen identification, morphological analysis, and risk assessment. It contains 210 SEM images distributed equally across three bacterial classes: Staphylococcus aureus, Streptococcus pneumoniae, and Vibrio cholerae.
Each class includes:
• 1 original SEM image taken from the Robert Koch Institute (RKI) electron microscopy image collection with proper attribution.
• 49 augmented images created using geometric and photometric transformations such as rotations and horizontal/vertical flips.
• 20 synthetic images generated using class-wise latent diffusion models, adding realistic variations while preserving biological structure.
The dataset is divided into training, validation, and test subsets to facilitate robust model development and evaluation. The training set primarily contains original and augmented images to capture class-specific morphological features, while the validation and test sets contain exclusively synthetic diffusion-generated images to evaluate model generalization, robustness, and performance on unseen, realistically varied data. This design ensures balanced representation across classes and maintains scientific transparency by clearly distinguishing between real, augmented, and synthetic images.
The final dataset structure is:
• Staphylococcus aureus – Train: 50, Validation: 12, Test: 8
• Streptococcus pneumoniae – Train: 50, Validation: 12, Test: 8
• Vibrio cholerae – Train: 50, Validation: 12, Test: 8
This carefully constructed dataset provides a reproducible, scientifically transparent, and defendable resource for deep learning, machine learning, and methodological research in SEM-based bacterial image classification. All original SEM images are credited to RKI, and all augmented and diffusion-generated images were created by the dataset authors exclusively for research purposes. By combining real, augmented, and diffusion-generated samples in a controlled, balanced, and low-data regime design, this dataset enables researchers to rigorously evaluate model performance, generalization, and the impact of synthetic data while preserving morphological integrity and scientific honesty.