Temporal ASTRA: Synthetic Evaluation and Hybrid CNN-BiLSTM Modeling for Calibration-Free Strabismus Detection

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Strabismus screening in pediatric and remote-care settings remains difficult because many existing methods depend on patient cooperation, individual calibration procedures, and static image capture, which are insufficient for detecting intermittent or transient ocular misalignment. The objective of this study is to introduce a calibration-free pre-screening approach that relies on temporal binocular behavior rather than absolute gaze measurements. We present Temporal ASTRA (Automatic Strabismus Tracking and Risk Assessment), a video-based framework that analyzes interocular disparity and its temporal evolution, including velocity and acceleration, from short binocular video segments. To address the limited availability of annotated clinical time-series data, a synthetic data generation process was developed to reproduce physiologically plausible normal and abnormal vergence patterns, such as gradual drift, intermittent phoria, and nystagmus-like oscillations. A hybrid convolutional neural network and bidirectional long short-term memory (CNN–BiLSTM) model with attention pooling was trained on the synthetic dataset and subsequently fine-tuned using real video recordings. The proposed system achieved 93.3% accuracy on held-out synthetic data and 90.9% accuracy with an AUC of 93.7% on real-world videos following synthetic pretraining. Evaluation on a clinical validation set of 24 videos yielded 100% sensitivity and 66.7% specificity at a high-sensitivity screening threshold. This study demonstrates that modeling temporal vergence dynamics provides a practical and robust basis for calibration-free, video-based strabismus pre-screening suitable for telemedicine and community-scale deployment.

This publication has 38 references indexed in Scilit: