目的·探究计算机辅助下自体荧光图像定量结果与口腔白斑病上皮异常增生等级的相关性。方法·纳入2016年4月—2024年1月于上海交通大学医学院附属第九人民医院口腔黏膜病科就诊的口腔白斑病患者357例。利用手持自体荧光仪器获取患者病损的自体荧光图像，将自体荧光图像转为灰度图像，获得量化指标。在Python中拟合有序多元Logistic回归模型，绘制累积概率图。将数据集划分训练集和测试集，生成决策树，调整不同的超参数，获得最佳的模型效果。计算准确度、精确度和F1分值。利用混淆矩阵对模型性能进行可视化呈现。结果·随着上皮异常增生程度的增加，相对色阶平均值呈现下降趋势。在上皮异常增生二分类中累积概率图不同类别曲线之间无重叠，在四分类中仅上皮重度异常增生与其他类别曲线有重叠，说明模型的区分能力较好。在二分类病理等级中，当训练集和测试集比例为4∶1、决策树最大深度为2时，准确度、精确度、F1分值可达到较高，分别为0.792、0.801和0.795。在四分类病理等级中，当训练集和测试集比例为9∶1、决策树最大深度为4时，准确度、精确度、F1分值可达到较高，分别为0.611、0.537和0.569。结论·口腔黏膜病专科医师可将计算机辅助下自体荧光图像定量结果作为参考依据，预测口腔白斑病患者上皮异常增生程度，监控患者癌变风险。

Objective ·To explore the correlation between the quantitative results of autofluorescence imaging under computer assistance and the grade of epithelial dysplasia in oral leukoplakia. Methods ·From April 2016 to January 2024, 357 patients with oral leukoplakia who visited the Department of Oral Mucosal Diseases at Shanghai Ninth People′s Hospital, Shanghai Jiao Tong University School of Medicine, were included. Autofluorescence images of the lesions were obtained using a handheld autofluorescence device. These images were converted to grayscale images to obtain quantitative metrics. An ordered multinomial Logistic regression model was fitted in Python, and cumulative probability plots were generated. The dataset was divided into training and testing sets, and a decision tree was generated. Different hyperparameters were adjusted to achieve optimal model performance. Accuracy, precision, and F1 scores were calculated. The model performance was visualized using a confusion matrix. Results ·As the degree of epithelial dysplasia increased, the relative mean color level showed a declining trend. In the binary classification of epithelial dysplasia, there was no overlap between the cumulative probability curves of different categories. In the four-category classification, only severe epithelial dysplasia overlapped with other category curves, indicating good discriminative ability of the model. In binary pathological grading, when the training and testing set ratio was 4∶1 and the maximum depth was 2, the accuracy, precision, and F1 scores were 0.792, 0.801, and 0.795, respectively. In the four-category pathological grading, when the training and testing set ratio was 9∶1 and the maximum depth was 4, the accuracy, precision, and F1 scores were 0.611, 0.537, and 0.569, respectively. Conclusion ·Computer-assisted quantitative analysis of autofluorescence images can be used by oral mucosal specialists as a reference to predict the degree of epithelial dysplasia in patients with oral leukoplakia and to monitor their risk of cancer.