Gürses Barış Oğuz, Özer Nezaket Ezgi, Bölükbaşı Gaye, İlhan Betul, Gözen Adar, Boyacıoğlu Hayal, Güneri Pelin
Department of Mechanical Engineering, Faculty of Engineering, Ege University, Bornova, Izmir, Türkiye.
Department of Oral & Maxillofacial Radiology, Faculty of Dentistry, Ege University, Bornova, İzmir, Türkiye.
Clin Oral Investig. 2025 Sep 2;29(9):431. doi: 10.1007/s00784-025-06478-z.
To evaluate the diagnostic potential of surface texture features extracted from clinical images in objectively differentiating benign from malignant oral lesions, and to validate classification performance of a Support Vector Machine (SVM) model using these features.
This study included 275 intraoral photographs of oral mucosal lesions with biopsy-confirmed diagnoses, sourced from both institutional archives and a public dataset. Lesion areas were manually annotated and converted into 3D surface plots to extract grayscale-based texture features. Eight statistical descriptors-mean, mode, median, variance, skewness, kurtosis, coefficient of variation (CoV), and entropy-were computed and normalized relative to adjacent healthy mucosa. Group differences were analyzed using MANOVA and effect size metrics (Cohen's d, eta squared). A support vector machine (SVM) with a Gaussian kernel was trained using five-fold cross-validation to classify lesions as benign or malignant based on the extracted features.
Statistical analysis revealed significant differences between benign and malignant groups for all features except skewness (p < 0.001). Entropy, kurtosis, and CoV showed the largest effect sizes, with entropy notably higher in malignant lesions and kurtosis higher in benign ones. The SVM model achieved a sensitivity of 99.2%, specificity of 81.4%, overall accuracy of 90.5%, and an AUC of 0.939, demonstrating high diagnostic performance in distinguishing malignant from benign oral mucosal lesions based on surface texture analysis.
Surface texture features, particularly entropy and kurtosis, offer promising diagnostic indicators for distinguishing malignant from benign lesions. SVM classifier demonstrated robust performance using these parameters.
This study highlights surface texture as an objective, underexplored diagnostic parameter. Integrating surface topography into clinical assessments and AI-based tools may enhance early detection and diagnostic accuracy in oral cancer screening.
评估从临床图像中提取的表面纹理特征在客观区分口腔良性与恶性病变方面的诊断潜力，并验证使用这些特征的支持向量机（SVM）模型的分类性能。
本研究纳入了275张口腔黏膜病变的口腔内照片，这些病变的诊断均经活检确认，照片来源包括机构档案和一个公共数据集。病变区域经手动标注并转换为三维表面图，以提取基于灰度的纹理特征。计算了八个统计描述符——均值、众数、中位数、方差、偏度、峰度、变异系数（CoV）和熵，并相对于相邻的健康黏膜进行归一化处理。使用多变量方差分析（MANOVA）和效应量指标（科恩d值、η²）分析组间差异。使用高斯核的支持向量机（SVM）通过五折交叉验证进行训练，以根据提取的特征将病变分类为良性或恶性。
统计分析显示，除偏度外，所有特征在良性和恶性组之间均存在显著差异（p < 0.001）。熵、峰度和CoV显示出最大的效应量，其中熵在恶性病变中显著更高，而峰度在良性病变中更高。SVM模型的灵敏度为99.2%，特异度为81.4%，总体准确率为90.5%，曲线下面积（AUC）为0.939，表明基于表面纹理分析在区分口腔恶性与良性黏膜病变方面具有较高的诊断性能。
表面纹理特征，特别是熵和峰度，为区分恶性与良性病变提供了有前景的诊断指标。SVM分类器使用这些参数表现出稳健的性能。
本研究强调表面纹理是一个客观的、尚未充分探索的诊断参数。将表面形貌纳入临床评估和基于人工智能的工具中，可能会提高口腔癌筛查的早期检测和诊断准确性。
