Deep learning application to hyphae and spores identification in fungal fluorescence images

This study explores the application of deep learning to fungal disease diagnosis, focusing on an automated detection system for hyphae and spores in clinical samples. This study employs a combination of the YOLOX and MobileNet V2 models to analyze fungal fluorescence images. The YOLOX model is used to identify individual fungal spores and hyphae, and the MobileNet V2 model is employed to identify fungal mycelium. Finally, their combination yields the results of the two analysis processes, providing positive or negative results for the entire sample set. The proposed dual-model framework is evaluated in terms of the precision, recall, F1-score, and Kappa metrics. For the YOLOX model, the precision is 85% for hyphae and 77% for spores, and for the MobileNet V2 model, the precision is 83%. The recall value of the YOLOX model is 90% for hyphae and 85% for spores, and that of the MobileNet V2 model is 100%. The agreement of the proposed dual-model framework with the doctors' evaluations in terms of precision, recall, and Kappa values is 92.5%, 99.3%, and 0.857, respectively. The high agreement value suggests the proposed dual-model framework's ability to identify fungal hyphae and spores in fluorescence images can reach the level of clinicians. With the help of the proposed framework, the time and labor of fungal diagnosis can be significantly saved.