Using deep learning methods to automatically interpret blood culture Gram stains

Abstract Background Gram staining of the smear prepared as soon as the growth signal is received in automated blood culture systems is very important in terms of providing the first critical information for the clinician to plan the appropriate treatment. However, microscopic interpretation of Gram-stained smears is one of the most time-intensive processes. At this stage, the use of deep learning techniques will be beneficial for us. Methods In the blood cultures sent to İzmir Bakırçay University Çiğli Training and Research Hospital Microbiology Laboratory during the project period, two smears with the same thickness were prepared with the same technique from those with positive growth signals. The smears were stained on a fully automated Gram staining device and digitized with a slide scanning and imaging device. After manual labeling of the micro-organisms in the images obtained, work was carried out on the training set using image processing and current deep learning techniques, and the analysis results were supported by the test set. Results We used the deep learning models xresnet50, resnet50, xresnext50, and mobilenetV3. The results indicate that it may be possible to develop a blood culture slide evaluation system using a deep learning model, particularly outside laboratory working hours. Conclusions Developing an automated system for Gram staining interpretation is crucial for ensuring uninterrupted laboratory operations both during and outside working hours. This will also contribute to antimicrobial stewardship by reducing the time it takes for a laboratory to issue its first report after a positive blood culture signal.