Deep learning for the identification of Candida spp. directly from blood culture gram stains from candidemia patients.

Candidemias caused by the yeasts formerly encompassed as Candida spp. require expedited identification to decide on the antifungal treatment and reduce mortality. Traditional methods rely on subcultures for diagnosis, with turnaround times of 72-96 h, or expensive equipment. Deep learning and convolutional neural networks (CNN) have shown high accuracy for image recognition in microbiology. We compared the accuracy of six CNNs (GoogLeNet, InceptionV3, AlexNet, ResNet18, ResNet50, and DenseNet161) to identify Candida spp. at the species level, with photographs obtained mainly from clinical blood cultures showing yeast structures in the Gram stain, which were identified as Candida spp. in the subculture. Images were obtained from January 2012 to May 2024 and stored in the image databank of two third-level teaching hospitals in Mexico City. We analyzed the five most frequent species from both centers' clinical samples and included simulated blood culture images from Candida auris (Candidozyma auris) and C. krusei (Pichia kudriavzevii) strains. After processing and segmentation, we loaded the CNNs with 531 whole photographs and 2804 patches. The CNN Densnet161, using a scan-based approach, showed higher accuracy identifying 87%, 99%, 94%, 100%, 89%, and 95% of the images containing C. albicans, C. auris, C. glabrata (Nakaseomyces glabrata), C. krusei (P. kudriavzevii), C. parapsilosis, and C. tropicalis, respectively. These results show that CNN image recognition can identify clinically relevant Candida spp. directly from positive Gram-stained smears, which may help make early decisions for antifungal treatment.