Photo credit: istock.com/jes2ufoto
Deep learning models using OCT imaging successfully predicted moderate to high accuracy SANS.
This study was published in the June 2025 issue of the American Journal of Ophthalmology. Researchers conducted retrospective studies to develop deep learning artificial intelligence (AI) models to predict spaceflight-related neuro-ocular syndromes (SANSs) using optical coherence tomography (OCT) images. Nerve Head (ONH).
They trained an AI deep learning model to predict the onset of SANS using the February 2nd dataset. Images of the pre-astronaut and in-flight images (flight data) and pre-bedrest and bederest from participants who received head-down tilt bed rest (HDTBR) as ground models (ground data). The dataset was split by participants for training and testing. The ResNet50-based model was trained individually with flight data, ground data, and combined datasets. Model performance was assessed using Preflight or Pre-Bedrest OCT images only. Prediction accuracy was measured using the receiver operating characteristics (ROC) region (AUC) under the curve. Class activation maps (CAMs) were generated to highlight important image regions that contribute to SANS predictions.
The results showed that the model trained with OCT data achieved a ROC AUC of 0.82 (95% CI: 0.54 – 1.0) for flight data and 0.67 (95% CI: 0.51 – 0.83) for ground HDTBR data. The model trained with HDTBR data reached an AUC of 0.71 (95% CI: 0.50 – 0.91) for ground data and 0.76 (95% CI: 0.51 – 0.91) for flight data. The combined model generated AUCs of 0.81 (95% CI: 0.53 – 0.95) for flight data and 0.72 (95% CI: 0.52 – 0.92) for ground data. The cam emphasized the nipple periphery Nerve Fibrous layers, retinal pigment epithelium, and anterior layer surfaces as major predictors.
Investigators concluded that the AI model showed moderate to high accuracy in predicting SANS from pre-lite OCT imaging, and that consistent performance across datasets supported HDTBR as a valid earth-based model for SANS.
sauce: ajo.com/article/S0002-9394(25)00298-3/Abstract
