Enhanced Vision Transformer with Local Feature Refinement for Fine-Grained Vocal Fold Lesion Classification

Accurate classification of vocal fold lesions is important for early clinical management, but remains challenging in laryngoscopic images because of high inter-class similarity, substantial intra-class variation, and highly localized discriminative features. To address these challenges, we proposed FocalViT, an enhanced Vision Transformer framework consisting of three components: a Dual Attention mechanism for channel and spatial feature refinement, a Local Feature Selection module for identifying discriminative pathological details, and a Structured Logit Regularization Loss for incorporating semantic relationships among lesion classes. On a dataset of 6560 laryngoscopic images covering carcinoma, keratosis, polyps, and normal controls, FocalViT achieved an overall accuracy of 90.7%, representing an 8.7-percentage-point improvement over the standard ViT baseline. For carcinoma, Recall increased from 76.9% to 92.4%, while Average Precision (AP) increased from 91.0% to 96.5%. These results indicate improved discrimination across different decision thresholds. Ablation studies showed that the three modules contributed individually and also improved performance when used together. Attention heatmaps further showed that the model mainly focused on lesion-related regions. These results suggest that FocalViT may serve as a computer-aided tool for laryngoscopic lesion assessment and provide additional support for the identification of high-risk cases.