International Journal of Science, Engineering and TechnologyAuthors- Priyanka Kelkar, Pranjal Aanjna, Preeti Anjana, Asst. Prof. Garima S. Tukra, Asst. Prof. Kailash Kumar Baraskar
Abstract-– Antimicrobial resistance (AMR) is a critical global health challenge, resulting in over 1.27 million deaths annually due to ineffective antibiotic treatments. Traditional AMR detection methods, including culture tests and PCR-based assays, are time-consuming and resource-intensive. In this study, we propose an automated AMR detection framework using Vision Transformers (ViTs) to classify blood culture images into resistant (R) and susceptible (S) categories. The ViT architecture effectively processes high-resolution blood culture images by leveraging multi-head attention mechanisms, enabling the model to capture fine-grained morphological differences associated with resistance patterns. The proposed model was evaluated on the PathAI AMR Image Dataset, containing 20,000 labeled blood culture images, achieving an accuracy of 94.2%, surpassing state-of-the-art CNN-based models by 6.7%. The model also achieved a precision of 92.8%, a recall of 95.4%, and an F1-score of 94.1%. Furthermore, Grad-CAM-based attention heatmaps provided interpretability, highlighting bacterial regions contributing to the resistance classification. The results demonstrate that ViT-based AMR detection offers superior accuracy, faster inference, and greater interpretability compared to traditional methods. This automated framework has the potential to significantly reduce diagnostic time in clinical settings, enabling faster and more precise antibiotic administration, thereby improving patient outcomes.
DOI: /10.61463/ijset.vol.13.issue2.406
