口腔白斑（OL）具有发展为口腔癌的内在倾向。患有上皮不典型增生（OED）的 OL 极有可能发生恶变；然而，常规的《牛津英语词典》评估具有侵入性且具有挑战性。本研究调查了深度学习 (DL) 模型是否可以使用口腔照片预测白斑患者的不典型增生概率。此外，我们还与临床医生的评分进行比较，评估了 DL 模型的性能，并为不典型增生评估提供决策支持。获得活检/组织病理学之前拍摄的白斑回顾性图像以构建 DL 模型 (n = 2,073)。组织病理学后的 OED 状态被用作所有图像的黄金标准。我们首先开发、微调并内部验证了一个具有 EfficientNet-B2 主干的 DL 架构，该架构输出 OED 的预测概率、OED 状态和感兴趣区域热图。然后，我们在地理验证之前在时间队列上测试了 DL 模型的性能。我们还根据人类评分者对《牛津英语词典》状态提供的意见，评估了模型在外部验证中的表现。绩效评估包括区分、校准和潜在净收益。 DL 模型在测试时取得了良好的 Brier 分数、曲线下面积和平衡精度 0.124 (0.079-0.169)、0.882 (0.838-0.926) 和 81.8% (76.5-87.1)，以及 0.146 (0.112-0.18)、0.828外部验证时分别为 (0.792-0.864) 和 76.4% (72.3-80.5)。此外，与对所有患者进行活检相比，该模型在 OED 评估期间选择 OL 患者进行活检/组织病理学时具有更高的潜在净收益。外部验证还表明，在根据口腔图像对白斑的 OED 状态进行分类时，DL 模型的准确度高于 92.3% (24/26) 的人类评估者（平衡准确度：54.8%-79.7%）。总体而言，基于照片的智能模型可以通过良好的校准和辨别来预测白斑的 OED 概率和状态，这显示出为选择患者进行活检/组织病理学、避免不必要的活检并协助患者自我监测提供决策支持的潜力。

Oral leukoplakia (OL) has an inherent disposition to develop oral cancer. OL with epithelial dysplasia (OED) is significantly likely to undergo malignant transformation; however, routine OED assessment is invasive and challenging. This study investigated whether a deep learning (DL) model can predict dysplasia probability among patients with leukoplakia using oral photographs. In addition, we assessed the performance of the DL model in comparison with clinicians' ratings and in providing decision support on dysplasia assessment. Retrospective images of leukoplakia taken before biopsy/histopathology were obtained to construct the DL model (n = 2,073). OED status following histopathology was used as the gold standard for all images. We first developed, fine-tuned, and internally validated a DL architecture with an EfficientNet-B2 backbone that outputs the predicted probability of OED, OED status, and regions-of-interest heat maps. Then, we tested the performance of the DL model on a temporal cohort before geographical validation. We also assessed the model's performance at external validation with opinions provided by human raters on OED status. Performance evaluation included discrimination, calibration, and potential net benefit. The DL model achieved good Brier scores, areas under the curve, and balanced accuracies of 0.124 (0.079-0.169), 0.882 (0.838-0.926), and 81.8% (76.5-87.1) at testing and 0.146 (0.112-0.18), 0.828 (0.792-0.864), and 76.4% (72.3-80.5) at external validation, respectively. In addition, the model had a higher potential net benefit in selecting patients with OL for biopsy/histopathology during OED assessment than when biopsies were performed for all patients. External validation also showed that the DL model had better accuracy than 92.3% (24/26) of human raters in classifying the OED status of leukoplakia from oral images (balanced accuracy: 54.8%-79.7%). Overall, the photograph-based intelligent model can predict OED probability and status in leukoplakia with good calibration and discrimination, which shows potential for decision support to select patients for biopsy/histopathology, obviate unnecessary biopsy, and assist in patient self-monitoring.