Enhancing ocular sign detection: AI-based strategic segmentation for improved accuracy and privacy protection

Accurate detection of ocular signs is essential for early diagnosis of eye diseases, but current AI approaches using facial or external ocular images include non-essential information, compromising performance and patient privacy. We conducted a multinational retrospective study of 2360 eyes from 1180 half-face images of thyroid eye disease patients across five racial groups from five hospitals in three countries. We developed a Dense Squeeze-and-Excitation Network (DSE-Net) to segment eyelid, conjunctiva, lacrimal caruncle, and eyeball, minimizing exposure and enhancing privacy. DSE-Net achieved Dice coefficient of 84.7%, 84.8%, 92.7%, and 95.1%, outperforming seven segmentation models. We then built SegmenView, employing LeNet, AlexNet, ResNet50, and VGGNet16 to detect eyelid edema, conjunctival erythema, caruncle or plica edema, and exophthalmos. SegmenView achieved internal Area Under the Curve (AUCs) of 71.09%, 80.81%, 90.07%, and 82.86%; external AUCs ranging 55.58%-84.29% across two test datasets, outperforming half-face and periocular models. We also compared SegmenView with four privacy-preserving methods, showing its superior ability to balance privacy protection with diagnostic accuracy. Additionally, visualizations based on Gradient-weighted Class Activation Mapping (Grad-CAM) further enhanced the model's interpretability. Our approach demonstrates high accuracy, generalizability, and potential for lightweight, privacy-preserving ocular sign detection.