Classification of the most common conditionally pathogenic microorganisms on SEM images with YOLO model

Abstract

A relevant and highly demanded modern medicine problem in many of its areas is the timely detection and recognition of pathogenic microorganisms and microbial communities in the patient’s tissues for the speedy prescription and correct use of medicines from mutually exclusive tactics. The transition to a new level in the speed of visualization of the samples’ contents taken and the accuracy of diagnostics is possible because of the use of lanthanide staining in combination with scanning electron microscopy to retrieve a series of high-resolution images with subsequent automatic labelling and classification of microbiological objects. This paper presents the results of using the YOLOv5 neural network model to detect 15 different most common opportunistic classes of bacteria in 380 images. As a result, a 71.5% average accuracy and 69.8% recall were achieved by using the YOLOv5 base model without freezing layers.