Bridging lab-to-clinic: microbiological screening via Swin-Ultra Transformer with transfer learning

Abstract

Bacterial infections are a critical global health issue, requiring rapid and precise pathogen identification for effective infection control. Traditional methods, such as culture and nucleic acid amplification, are often slow and lack sensitivity. Raman spectroscopy combing with deep learning has been a powerful technique for microbial identification. However, limitations such as bacterial physiological states, genetic variation, interference from biological materials, and differences in laboratory conditions make its practical application still challenging. This study introduces a feature-enhanced dual-attention pathway Shifted Window-Ultra (Swin-Ultra) Transformer architecture, integrated with deep transfer learning, to address challenges like bacterial physiological states, genetic variation, and laboratory condition discrepancies. A Bacterial Pre-trained Transformer (BPT) was developed using the Bacteria-ID database, achieving excellent classification performance, i.e., 98.26 % accuracy. Fine-tuning with clinical datasets yielded accuracies of 99.80 % for bacterial pathogens and 98.53 % for Cryptococcus genotypes. This approach, bridges laboratory models and clinical applications, enhancing unknown pathogen identification, infection control, and public health surveillance, with significant potential to improve patient outcomes.