Abstract
Malignant and premalignant ocular surface tumors (OSTs) can be sight-threatening or even life-threatening if not diagnosed and treated promptly. Artificial intelligence holds great promise for the early detection of these diseases. However, training traditional convolutional neural networks (CNNs) for this task presents challenges due to the lack of large, well-annotated datasets containing OST images labeled according to histopathological results. Here, we introduce the ocular surface pretrained model (OSPM), a domain-specific pretrained model designed to address the scarcity of labeled data. OSPM is constructed utilizing self-supervised learning on approximately 0.76 million unlabeled ocular surface images from 10 clinical centers across China and can be readily adapted to the OST classification task. We then develop and evaluate an OSPM-enhanced classification model (OECM) using 1,455 OST images labeled with histopathological diagnoses to differentiate between malignant, premalignant, and benign OSTs. OECM achieves excellent performance with AUROCs ranging from 0.891 to 0.993 on internal, external, and prospective test datasets, significantly outperforming the traditional CNN models. OECM demonstrated performance comparable to that of senior ophthalmologists and increased the diagnostic accuracy of junior ophthalmologists. Greater label efficiency was observed in OECM compared to CNN models. Our proposed model has high potential to enhance the early detection and treatment of malignant and premalignant OSTs, thereby reducing cancer-related mortality and optimizing functional outcomes.