Deep Learning Based on DeepLabv3+ for Automated Identification, Segmentation, and Quantification of Fungi in Clinical Fluorescence Microscopy Images.

BACKGROUND: Fungal fluorescence microscopy using Calcofluor White is a rapid and sensitive diagnostic tool for fungal infections. However, manual microscopic examination is limited by high subjectivity, observer fatigue, poor reproducibility, and a heavy reliance on professional expertise-particularly in high-throughput clinical screening. OBJECTIVES: To develop and validate a DeepLabv3+-based deep learning model for the automated identification, precise segmentation, and quantification of fungal hyphae and spores in clinical fluorescence microscopy images. METHODS: A total of 60,484 fungal fluorescence images were collected from 1863 clinical specimens across seven medical centers. Of these, 38,494 images were used for model training, while 2561 independent images were selected for clinical validation. Blinded manual microscopy by expert technicians served as the gold standard. The model's diagnostic performance, stability, and agreement were evaluated. RESULTS: The model demonstrated robust diagnostic performance, with an accuracy of 90.32%, precision of 91.58%, recall (sensitivity) of 92.42%, and specificity of 88.39%. The Youden index was 0.8081, with an area under the curve (AUC) of 0.91, indicating substantial agreement (Kappa = 0.798). The system exhibited excellent stability across repeated tests and automatically generated quantitative counts of hyphae and spores. CONCLUSIONS: This DeepLabv3+-based deep learning system provides an accurate, stable, and high-throughput approach for clinical fungal fluorescence microscopy. It can standardize diagnostic procedures and improve efficiency for diagnosing superficial and suspected invasive fungal infections, serving as a reliable auxiliary diagnostic tool.