Artificial intelligence for keratosis characterization and identification of lichenoid lesions in histological samples of oral leukoplakia

Comprehensive Summary

In this paper, the application of artificial intelligence (AI) and computational pathology techniques to automatically assess the histology of oral leukoplakia (OL) and oral lichenoid lesions (OLL) was investigated. OL and OLL are examples of potentially malignant disorders linked to the development of oral squamous cell carcinoma (OSCC). In this study, 240 hematoxylin and eosin–stained slides from 192 patients were analyzed retrospectively using deep learning-based algorithms for tissue segmentation, morphometry, and classification of lesions. The nnU-Net framework provided accurate tissue segmentation of epithelium, subepithelium, keratin, and nuclei, reaching Dice scores above 0.92 for the majority of tissues. Keratin thickness, epithelial histology, and nuclei density were quantified, enabling identification of keratinization types (orthokeratosis and parakeratosis), with an accuracy close to 92%. Furthermore, a Tiny Vision Transformer (TinyViT) was able to classify OL and OLL with lesion-level accuracy reaching nearly 93%.

Outcomes and Implications

The significance of this study from a medical point of view lies in the similarities between oral leukoplakia and oral lichenoid lesions in their histopathological presentation. Due to the fact that each lesion poses a different risk for developing oral squamous cell carcinoma, proper diagnosis plays a vital role in identifying appropriate surveillance measures and clinical guidelines. The proposed artificial intelligence-assisted pathology workflow offers an objective and reproducible system that will allow the accurate assessment of keratinization patterns and classification of oral lesions. Using automated morphometric analysis could potentially facilitate the identification of any epithelial changes that would be difficult to detect using conventional microscopy alone. Early and consistent identification of high-risk lesions could contribute to patient risk stratification, biopsy evaluation, and surveillance for malignant transformation. Another potential implication of this study could involve the use of transformer-based deep learning architectures in digital pathology tools that could help pathologists deliver precise diagnoses in areas with a lack of oral pathology expertise.