Saldivia-Siracusa Cristina, Carlos de Souza Eduardo Santos, Barros da Silva Arnaldo Vitor, Damaceno Araújo Anna Luíza, Pedroso Caíque Mariano, Aparecida da Silva Tarcília, Pereira Sant'Ana Maria Sissa, Fonseca Felipe Paiva, Rebelo Pontes Hélder Antônio, Quiles Marcos G, Lopes Marcio Ajudarte, Vargas Pablo Agustin, Khurram Syed Ali, Pearson Alexander T, Lingen Mark W, Kowalski Luiz Paulo, Hunter Keith D, Ponce de Leon Ferreira de Carvalho André Carlos, Santos-Silva Alan Roger
Departamento de Diagnóstico Oral, Faculdade de Odontologia de Piracicaba, Universidade Estadual de Campinas, Piracicaba, São Paulo, Brazil.
Institute of Mathematics and Computer Sciences, University of São Paulo, São Carlos, São Paulo, Brazil.
Lancet Reg Health Am. 2025 May 29;47:101138. doi: 10.1016/j.lana.2025.101138. eCollection 2025 Jul.
Artificial Intelligence (AI) models hold promise as useful tools in healthcare practice. We aimed to develop and assess AI models for automatic classification of oral potentially malignant disorders (OPMD) and oral squamous cell carcinoma (OSCC) clinical images through a Deep Learning (DL) approach, and to explore explainability using Gradient-weighted Class Activation Mapping (Grad-CAM).
This study assessed a dataset of 778 clinical images of OPMD and OSCC, divided into training, model optimization, and internal testing subsets with an 8:1:1 proportion. Transfer learning strategies were applied to pre-train 8 convolutional neural networks (CNN). Performance was evaluated by mean accuracy, precision, recall, specificity, F1-score and area under the receiver operating characteristic (AUROC) values. Grad-CAM qualitative appraisal was performed to assess explainability.
ConvNeXt and MobileNet CNNs showed the best performance. Transfer learning strategies enhanced performance for both algorithms, and the greatest model achieved mean accuracy, precision, recall, F1-score and AUROC of 0.799, 0.837, 0.756, 0.794 and 0.863 during internal testing, respectively. MobileNet displayed the lowest computational cost. Grad-CAM analysis demonstrated discrepancies between the best-performing model and the highest explainability model.
ConvNeXt and MobileNet DL models accurately distinguished OSCC from OPMD in clinical photographs taken with different types of image-capture devices. Grad-CAM proved to be an outstanding tool to improve performance interpretation. Obtained results suggest that the adoption of DL models in healthcare could aid in diagnostic assistance and decision-making during clinical practice.
This work was supported by FAPESP (2022/13069-8, 2022/07276-0, 2021/14585-7 and 2024/20694-1), CAPES, CNPq (307604/2023-3) and FAPEMIG.
人工智能（AI）模型有望成为医疗保健实践中的有用工具。我们旨在通过深度学习（DL）方法开发和评估用于自动分类口腔潜在恶性疾病（OPMD）和口腔鳞状细胞癌（OSCC）临床图像的AI模型，并使用梯度加权类激活映射（Grad-CAM）探索可解释性。
本研究评估了一个包含778张OPMD和OSCC临床图像的数据集，按照8:1:1的比例分为训练集、模型优化集和内部测试集。应用迁移学习策略对8个卷积神经网络（CNN）进行预训练。通过平均准确率、精确率、召回率、特异性、F1分数和受试者工作特征曲线下面积（AUROC）值评估性能。进行Grad-CAM定性评估以评估可解释性。
ConvNeXt和MobileNet CNN表现最佳。迁移学习策略提高了两种算法的性能，在内部测试中，性能最佳的模型的平均准确率、精确率、召回率、F1分数和AUROC分别达到0.799、0.837、0.756、0.794和0.863。MobileNet的计算成本最低。Grad-CAM分析表明性能最佳的模型与可解释性最高的模型之间存在差异。
ConvNeXt和MobileNet DL模型在使用不同类型图像采集设备拍摄的临床照片中准确区分了OSCC和OPMD。Grad-CAM被证明是提高性能解释的出色工具。获得的结果表明，在医疗保健中采用DL模型有助于临床实践中的诊断辅助和决策。
本研究得到了圣保罗研究基金会（FAPESP，项目编号2022/13069-8、2022/07276-0、2021/14585-7和2024/20694-1）、巴西高等教育人员素质提升协调办公室（CAPES）、巴西国家科学技术发展委员会（CNPq，项目编号307604/2023-3）和米纳斯吉拉斯州研究资助基金会（FAPEMIG）的支持。
