AI-Assisted Three-Class Oral Histopathology: A Reader Study

INTRODUCTION: Oral squamous cell carcinoma (OSCC) and oral leukoplakia (OLK) impose a growing diagnostic burden on oral pathology services, particularly in primary care. The present study developed a three-class deep learning model to discriminate Normal, OLK and OSCC on mixed-magnification H&E histopathology patches and evaluated its adjunctive value in a controlled human-AI reader study. METHODS: Three publicly available oral histopathology datasets were harmonised, yielding 25,576 patches partitioned at 8:1:1 into training, validation and test sets, with 300 balanced patches reserved for reader assessment. VGG16-BR was built on a VGG16-BN backbone with squeeze-and-excitation and global average pooling, and was compared with ResNet50 and ViT-Base. Six oral pathologists (3 junior, 3 senior) completed the reader task without and then with AI assistance, separated by a 2-week washout. RESULTS: VGG16-BR achieved test-set accuracy 0.9065, macro-F1 0.9086 and macro-AUC 0.9628, with the lowest OSCC-to-Normal misclassification rate (5.1%). AI assistance increased mean accuracy from 0.767 to 0.857 in junior readers and from 0.850 to 0.894 in senior readers, raised Fleiss' κ and shortened reading time. The mixed model showed a significant main effect of AI assistance (P < .001), with no significant interaction between AI assistance and reader experience (P = .265). CONCLUSION: A lightweight three-class deep learning framework improved diagnostic accuracy, inter-reader agreement and efficiency in oral histopathology reading. CLINICAL RELEVANCE: Positioned as an adjunct rather than a replacement for pathologists, the model provides three-class diagnostic support that may strengthen the early recognition of OLK and OSCC in resource-limited primary care settings.