Application of few-shot learning and transfer learning based on YOLOv6 in the recognition of bacteria in sputum M-ROSE.

Background: Rapid pathogen identification is essential for guiding timely and appropriate antimicrobial therapy in severe pulmonary infections. Microbiological rapid on-site evaluation (M-ROSE) can provide preliminary etiological information at the bedside, but its interpretation is labor intensive and highly dependent on experienced clinicians or microbiologists. This study aimed to develop and evaluate a deep learning framework for bacterial identification in sputum M-ROSE smears under limited annotated data conditions. Methods: A total of 161 Gram-stained sputum M-ROSE images containing Acinetobacter baumannii, Klebsiella pneumoniae, and Pseudomonas aeruginosa were collected using a digital scanner. Thirty annotated images were used to construct Dataset N, which was divided into a training set and a validation set at a ratio of 8:2. A deep learning model based on the YOLOv6 architecture was developed using transfer learning and few-shot learning. In addition to conventional data augmentation, two customized strategies-rotation cutting and image fusion-were introduced to enhance the detection of extremely small bacterial targets. The remaining 131 images were used as an independent testing set. Model performance was evaluated using recall, precision, F1 score, mean average precision (mAP), accuracy, and diagnostic time. Results: The best-performing model achieved 89.31% recall, 84.23% precision, an F1 score of 0.8670, and an mAP of 0.817 on the validation set. On the independent test set, the model achieved 93.89% accuracy. In addition, the model required substantially less time for image interpretation than the participating clinicians. Conclusion: The proposed YOLOv6-based framework showed good performance for bacterial identification in sputum M-ROSE smears under limited annotated data conditions. These findings support the feasibility of applying data-efficient deep learning strategies to real-world clinical microbiological images and suggest potential utility in rapid microbiological diagnosis.