Anisotropy-Aware Cellular Automaton Framework for AI-Powered Pterygium Screening

Purpose: The purpose of this study was to explore a weakly-labeled, positive-sample-free auxiliary diagnostic method for pterygium, providing an effective screening tool for primary healthcare institutions and self-examination. Methods: An ocular surface anatomic atlas was initially constructed through semantic segmentation, and an improved cellular automaton method was used to simulate the characteristic anisotropic growth pattern of pterygium, thereby generating high-quality training samples. Results: The proposed method effectively overcomes the dependency on professionally annotated data in traditional artificial intelligence (AI) diagnosis, with model performance surpassing conventional training approaches, enabling accurate identification of pterygium lesions. Conclusions: The pathomimetic computational framework presented in this study offers an innovative approach to automatically generate labeled pterygium cases, facilitating auxiliary-diagnostic AI model for early screening. It is particularly suitable for primary healthcare settings with limited ophthalmic resources and holds significant potential for reducing the risk of vision impairment. Translational Relevance: This study bridges a weakly-supervised, positive-sample-free approach for pterygium detection and clinical care by developing an AI-powered screening tool that enables early identification in resource-limited primary healthcare settings, facilitating timely intervention and preventing vision impairment.