Majdy Idrees1, Camile S Farah2,3,4, Kate Shearston1, Omar Kujan1. 1. UWA Dental School, The University of Western Australia, Nedlands, WA, Australia. 2. Australian Centre for Oral Oncology Research and Education, Nedlands, WA, Australia. 3. Oral, Maxillofacial and Dental Surgery, Fiona Stanley Hospital, Murdoch, WA, Australia. 4. Anatomical Pathology, Australian Clinical Labs, Subiaco, WA, Australia.
Abstract
BACKGROUND: Oral lichen planus is a relatively common oral disorder which shares clinical and histopathological features with other lichenoid lesions, leading to considerable inter-observer disagreement. This negatively impacts understanding of the pathogenesis and malignant transformation potential of this condition. METHODS: Artificial intelligence was employed to create a machine-learning artificial neural network to identify and quantify mononuclear cells and granulocytes within the inflammatory infiltrates in digitized hematoxylin and eosin microscopic slides. Twenty-four regions of interest were extracted from oral lichen planus cases for learning purposes and validated on a retrospective cohort of 130 cases. All cases were related to patients with confirmed diagnoses of oral lichen planus, oral lichenoid lesions, or oral epithelial dysplasia with lichenoid host response. RESULTS: The number of inflammatory cells was statistically significantly higher in oral lichen planus compared to oral lichenoid lesions or oral epithelial dysplasia with lichenoid host response (p <.0005). The proposed machine-learning method was reliably capable of detecting oral lichen planus cases based on the number of inflammatory cells and the number of mononuclear cells with an area under the curve of 0.982 and 0.988, respectively. Identifying a cut-off point between oral lichen planus and other lichenoid conditions based on the number of mononuclear cells showed a sensitivity of 100% and an accuracy of 94.62%. CONCLUSION: Artificial intelligence has shown promising outcomes and provides a robust approach to enhance the accuracy of anatomical pathologists in accurately diagnosing oral lichen planus using features of disease pathogenesis. This article is protected by copyright. All rights reserved.
BACKGROUND: Oral lichen planus is a relatively common oral disorder which shares clinical and histopathological features with other lichenoid lesions, leading to considerable inter-observer disagreement. This negatively impacts understanding of the pathogenesis and malignant transformation potential of this condition. METHODS: Artificial intelligence was employed to create a machine-learning artificial neural network to identify and quantify mononuclear cells and granulocytes within the inflammatory infiltrates in digitized hematoxylin and eosin microscopic slides. Twenty-four regions of interest were extracted from oral lichen planus cases for learning purposes and validated on a retrospective cohort of 130 cases. All cases were related to patients with confirmed diagnoses of oral lichen planus, oral lichenoid lesions, or oral epithelial dysplasia with lichenoid host response. RESULTS: The number of inflammatory cells was statistically significantly higher in oral lichen planus compared to oral lichenoid lesions or oral epithelial dysplasia with lichenoid host response (p <.0005). The proposed machine-learning method was reliably capable of detecting oral lichen planus cases based on the number of inflammatory cells and the number of mononuclear cells with an area under the curve of 0.982 and 0.988, respectively. Identifying a cut-off point between oral lichen planus and other lichenoid conditions based on the number of mononuclear cells showed a sensitivity of 100% and an accuracy of 94.62%. CONCLUSION: Artificial intelligence has shown promising outcomes and provides a robust approach to enhance the accuracy of anatomical pathologists in accurately diagnosing oral lichen planus using features of disease pathogenesis. This article is protected by copyright. All rights reserved.