Attention-Guided Hybrid Deep Learning Framework for Automated Oral Lichen Planus Classification

Oral lichen planus (OLP) is a chronic inflammatory disorder classified as an oral potentially malignant disorder. Accurate clinical diagnosis remains challenging due to lesion heterogeneity and substantial overlap with other oral mucosal diseases. This study developed an explainable artificial intelligence framework for automated OLP classification using complementary deep feature representations and hybrid feature optimization. A dataset comprising 1,345 clinical images (747 OLP and 598 non-OLP) was divided into training (n = 967), validation (n = 189), and test (n = 189) subsets using stratified sampling. Images were resized to 224 × 224 pixels, and geometric data augmentation was applied only to the training set, expanding it to 10,000 images while preventing data leakage. Two attention-guided convolutional neural networks, LichenNet-CBAM and Inception-LichenNet-CBAM, were independently trained to extract complementary 512-dimensional deep feature representations. The resulting 1,024-dimensional fused feature space was optimized using a hybrid Grey Wolf Optimizer with Mutation-based Local Search (GWO-MLS) prior to evaluation across multiple machine learning classifiers. Model interpretability was enhanced using Grad-CAM (Gradient-weighted Class Activation Mapping), feature importance analysis, and confidence calibration. On the independent test dataset, Naïve Bayes and AdaBoost achieved the highest classification accuracy of 94.71%, whereas Random Forest produced the highest discriminative performance (receiver operating characteristic area under the curve, ROC-AUC = 0.981). The Naïve Bayes classifier correctly identified 99 of 105 OLP lesions and 80 of 84 non-OLP lesions. Statistical analysis demonstrated overlapping 95% confidence intervals and no significant performance differences between the best-performing classifier and alternative models. The proposed framework provides accurate and interpretable automated classification of OLP by integrating complementary deep feature extraction, hybrid feature optimization, and explainable artificial intelligence. Further validation using large multicenter cohorts is warranted before clinical implementation.