Image-Based Diagnostic Performance of LLMs vs CNNs for Oral Lichen Planus: Example-Guided and Differential Diagnosis

Abstract

Introduction and aims

The overlapping characteristics of oral lichen planus (OLP), a chronic oral mucosal inflammatory condition, with those of other oral lesions, present diagnostic challenges. Large language models (LLMs) with integrated computer-vision capabilities and convolutional neural networks (CNNs) constitute an alternative diagnostic modality. We evaluated the ability of seven LLMs, including both proprietary and open-source models, to detect OLP from intraoral images and generate differential diagnoses.

Methods

Using a dataset with 1,142 clinical photographs of histopathologically confirmed OLP, non-OLP lesions, and normal mucosa. The LLMs were tested using three experimental designs: zero-shot recognition, example-guided recognition, and differential diagnosis. Performance was measured using accuracy, precision, recall, F1-score, and discounted cumulative gain (DCG). Furthermore, the performance of LLMs was compared with three previously published CNN-based models for OLP detection on a subset of 110 photographs, which were previously used to test the CNN models.

Results

Gemini 1.5 Pro and Flash demonstrated the highest accuracy (69.69%) in zero-shot recognition, whereas GPT-4o ranked first in the F1 score (76.10%). With example-guided prompts, which improved consistency and reduced refusal rates, Gemini 1.5 Flash achieved the highest accuracy (80.53%) and F1-score (84.54%); however, Claude 3.5 Sonnet achieved the highest DCG score of 0.63. Although the proprietary models generally excelled, the open-source Llama model demonstrated notable strengths in ranking relevant diagnoses despite moderate performance in detection tasks. All LLMs were outperformed by the CNN models.

Conclusion

The seven evaluated LLMs lack sufficient performance for clinical use. CNNs trained to detect OLP outperformed the LLMs tested in this study.