Accurate and Interpretable Oral Lichen Planus Classification Using a Hybrid Framework with Attention-Guided Feature Fusion and Hybrid Optimisation

This study utilised a dataset of 1,346 clinical photographs comprising 747 OLP and 599 non-OLP cases collected from the College of Dental Medicine, Rangsit University. To address class imbalance and improve model generalisation, we applied targeted augmentation, expanding the dataset to 8,000 balanced images. The dataset was stratified into training, validation and testing subsets using an 80:10:10 split. Develop two custom convolutional neural networks (CNNs), LichenNet and Inception-LichenNet, to capture both fine-grained and high-level features. Deep features were extracted from both models and integrated using an attention-guided feature fusion strategy. To further optimise efficiency and reduce redundancy, we applied a hybrid feature selection (GWO + See Tree) optimisation. For clinical interpretability, Grad-CAM was used to generate class-discriminative visual explanations. All experiments were conducted on Kaggle GPU environments using 10-fold cross-validation to ensure reliable evaluation

本研究使用了从兰实大学牙医学院采集的1346张临床照片数据集，其中包含747例口腔扁平苔藓（Oral Lichen Planus，简称OLP）病例与599例非OLP病例。为解决类别不平衡问题并提升模型泛化性能，本研究采用针对性数据增强策略，将数据集扩充至8000张平衡样本。该数据集以80:10:10的比例被分层划分为训练集、验证集与测试集。本研究构建了两款自定义卷积神经网络（Convolutional Neural Networks，简称CNN）：LichenNet与Inception-LichenNet，以同时捕获细粒度特征与高层语义特征。从两款模型中提取深度特征，并采用注意力引导的特征融合策略完成特征整合。为进一步优化模型效率并降低特征冗余，本研究应用了混合特征选择（GWO + See Tree）优化方案。为实现临床可解释性，本研究使用梯度类激活映射（Gradient-weighted Class Activation Mapping，简称Grad-CAM）生成类别判别性可视化解释。所有实验均在Kaggle GPU环境中开展，并采用10折交叉验证以确保评估结果的可靠性。