Accurate prognostic stratification of oral leukoplakia (OLK) with risk of malignant transformation into oral squamous cell carcinoma is crucial. We developed an objective and powerful pathomics-based model for the prediction of malignant transformation in OLK using hematoxylin and eosin (H&E)-stained images. In total, 759 H&E-stained images from multicenter cohorts were included. A training set (n = 489), validation set (n = 196), and testing set (n = 74) were used for model development. Four deep learning methods were used to train and validate the model constructed using H&E-stained images. Pathomics features generated through deep learning combined with machine learning algorithms were used to develop a pathomics-based model. Immunohistochemical staining of Ki67, p53, and PD-L1 was used to interpret the black box of the model. Pathomics-based models predicted the malignant transformation of OLK (validation set area under curve [AUC], 0.899; testing set AUC, 0.813) and significantly identified high-risk and low-risk populations. The prediction performance of malignant transformation from dysplasia grading (validation set AUC, 0.743) was lower than that of the pathomics-based model. The expressions of Ki67, p53, and PD-L1 were correlated with various pathomics features. The pathomics-based model accurately predicted the malignant transformation of OLK and may be useful for the objective and rapid assessment of the prognosis of patients with OLK.
Keywords: deep learning; histopathology; malignant transformation; oral leukoplakia; oral squamous cell carcinoma.
Copyright © 2023 United States & Canadian Academy of Pathology. Published by Elsevier Inc. All rights reserved.