Determination of Fungiform Papilla Number Using Deep Learning Methods

This study aimed to develop a deep learning-based method for the automatic detection and counting of fungiform papillae (FP) on the dorsal surface of the human tongue. FP density and morphology may serve as biomarkers for taste function and systemic disease diagnosis. Manual counting is time-consuming and subjective; therefore, an objective and reproducible artificial intelligence (AI) method was designed to provide a reliable quantitative assessment. A deeplearning object detection model was constructed using the Ultralytics YOLOv11 architecture. A dataset of 177 high-resolution toluidine blue-stained tongue images was manually annotated and dividedin to training, validation, and test sets. Three-foldnestedcross-validation was employed for hyperparameter optimization. Transfer learning was applied by freezing 22 backbone layers, and the detection heads were trained using tuned learning rates and decay factors. Early stopping was used to prevent overfitting. Model performance was evaluated on the independent test set. The model achieved 0.678 precision, 0.740 recall, and 0.707 F1 score, reflecting balanced detection performance. Compared with existing studies, our model demonstrated improved generalization and robustness. The mean absolute error (37.52; 19.48% of the true mean) and root mean square error (43.83) indicated reliable counting accuracy given the natural variability of FP counts (192.56 ± 63.14). The proposed YOLOv11-based model provides a fast, accurate, and reproducible alternative to manual FP counting. This approach may support large-scale clinical and research applications where FP analysis serves as a potential biomarker of health status.