TriGWONet a lightweight multibranch convolutional neural network using gray wolf optimization for accurate oral cancer image classification

Oral cancer remains a critical global health burden, with survival rates heavily dependent on early and accurate diagnosis. In recent years, deep learning-based medical image analysis has emerged as a promising tool for automated detection; however, existing models often suffer from excessive architectural complexity, limited accuracy, and inadequate generalization to clinical data. To address these limitations, we propose a novel lightweight multi-branch convolutional neural network (CNN) architecture integrated with Transformer-inspired design principles and optimized through Gray Wolf Optimization (GWO). The model comprises three parallel feature extraction branches, each incorporating SeparableConv2D layers, Batch Normalization, ReLU activation, and Global Max Pooling, followed by dense layers for hierarchical representation learning. GWO is employed to dynamically optimize critical hyperparameters—including the number of filters, dense layer neurons, and dropout rates—resulting in a compact and highly performant configuration.Performance is assessed on the OCI dataset (240 images; lips and tongue; cancerous vs. non-cancerous) with matched input size, augmentations, and tuning procedures for contemporary lightweight baselines (MobileViT, ConvNeXt, CoAtNet). The proposed TriGWONet model achieved a classification accuracy of 99.64%, sensitivity of 99.60%, specificity of 99.70%, precision of 99.50%, and an ROC AUC of 1.00. Comparative analysis demonstrates that our method significantly outperforms state-of-the-art models, including MobileViT, ConvNeXt, and CoAtNet. The superior accuracy and minimal computational footprint of the proposed TriGWONet model underscore its potential for real-time clinical deployment and pave the way for broader applications in medical image-based disease classification.