Boundary-aware segmentation of Oral Lichen Planus using parameter-efficient adaptation of DINOv2

Oral lichen planus (OLP) is a common chronic inflammatory disease of the oral mucosa that causes recurrent pain and discomfort, affecting patients’ quality of life. Accurate diagnosis is essential, given its potential progression to oral cancer. Segmentation masks allow the extraction of regions of interest for oral lesion diagnosis and often outperform full-image AI analysis, while precise lesion delineation remains crucial for assessing lesion extent, long-term monitoring, and clinical decision-making. However, OLP lesion segmentation remains challenging because of diffuse lesion boundaries and the low contrast between affected areas and surrounding healthy mucosa. In this study, we investigate the parameter-efficient adaptation of a foundation vision transformer, DINOv2, for medical image segmentation by combining low-rank adaptation (LoRA) with a DeepLabV3 decoder. A boundary-aware optimization integrating Dice, Cross-Entropy and differentiable Hausdorff losses is used to improve contour precision. The study was conducted on 668 OLP photographs, pixel-level annotated by a single oral medicine specialist. Evaluation included pixel-wise (Dice, IoU), boundary (HD95) and object-level metrics (TP/FP/FN, precision, recall, F1). Results were compared with established baselines. The proposed model achieved a Dice score of 0.8401, an IoU of 0.7734, and an HD95 of 62.68 px, outperforming all supervised CNN and Transformer baselines; for example, it improved over SegFormer MiT-B2 by +0.0283 Dice, +0.0305 IoU, and −8.12 px HD95. To assess robustness in low-data settings, we trained the model on subsets of different sizes and found that it consistently maintained its advantage. These results highlight the importance of calibrating boundary-sensitive loss components and object-level thresholds according to clinical priorities.