Deep-learning application for identifying histological features of epithelial dysplasia of tongue

Abstract

ObjectiveThis study aimed to evaluate the potential of artificial intelligence (AI) methods for histopathological grading of tongue dysplasia. MethodsA dataset comprising 203 digitized whole-slide images (WSIs) was constructed at 200× magnification from surgical specimens of tongue dysplasia and early cancers. Transfer learning was performed using this dataset with the convolutional neural network Inception-v3. Three experiments were conducted to (1) evaluate the potential of Inception-v3 by developing an AI system for epithelium classification and four AI systems for dysplasia grading (DysAI-1s), (2) compare the prediction performance of four DysAI-1s with that of humans, and (3) augment datasets to improve the accuracy of the developed AI system for dysplasia grading. ResultsThe area under the receiver operating characteristic curve (AUC) for epithelium classification was 0.996, while the average AUC scores of the four DysAI-1s were 0.921, 0.919, 0.867, and 0.774. In comparison, each AI system exhibited a moderate to substantial match (kappa, κ = 0.55−0.78) with the corresponding human predictions. The mean absolute errors (MAEs) showed a similar trend; MAEs of the DysAI-1s were 0.308, 0.403, 0.422, and 0.622. The final AI system (DysAI-2s) trained with larger datasets for dysplasia grading achieved an average κ-value of 0.81 despite different data-subset distribution ratios. Most MAEs of DysAI-2s showed better values than those of DysAI-1.1 (P < .001). ConclusionA properly trained AI system has the potential to improve the accuracy of oral dysplasia grading on WSIs.