A Clinical Bacterial Dataset for Deep Learning in Microbiological Rapid On-Site Evaluation

Area of Science:

• Medical Microbiology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Microbiological Rapid On-Site Evaluation (M-ROSE) aids pulmonary infectious disease diagnosis via microscopy.
• Automated pathogen identification is crucial for enhancing M-ROSE efficiency and accuracy.
• Existing deep learning studies often lack clinical data and focus on cultured bacteria.

Purpose of the Study:

• To develop and validate deep learning algorithms for automated bacterial identification from clinical specimens.
• To create a comprehensive dataset of Gram-stained bacteria from lower respiratory tract infections for research.
• To benchmark detection and segmentation networks for clinical M-ROSE applications.

Main Methods:

• Collected Gram-stained bacteria images from 2018-2022 using M-ROSE from patients with lung infections.
• Created a dataset of 1705 desensitized images (4,912 × 3,684 pixels).
• Manually labeled and differentiated 4,833 cocci and 6,991 bacilli (Gram-negative and Gram-positive).
• Applied detection and segmentation networks for benchmark testing.

Main Results:

• A novel, large-scale dataset of clinical bacterial images was established.
• The dataset includes manually labeled cocci and bacilli with Gram staining information.
• Benchmark testing of deep learning networks was performed on the clinical dataset.

Conclusions:

• The provided dataset and benchmark algorithms can advance automated bacterial identification in clinical settings.
• This work addresses the gap in clinical data for deep learning in M-ROSE.
• The findings support the development of AI-driven tools for faster and more accurate infectious disease diagnosis.