AI Predicts Myopia Progression, not by “reading a crystal ball,” but by analyzing massive amounts of eye data to find hidden patterns.👁️
AI Predicts Myopia: Two Different Tasks Current AI performs very differently on predicting onset versus predicting progression rate :· Predicting Onset (Future Myopia): Performs excellently. AI can precisely capture spatial features in the fundus through ResNet34 and track changes over time with LSTM, achieving AUC above 0.94 and even above 0.98 for high myopia .· Predicting Progression (How Fast It Worsens): Currently a global challenge. Because myopia progression is nonlinear, past progression history has limited predictive value for the future .
AI often excels at predicting “will myopia develop” while struggling with “how much will it increase next year” .💡
How AI “Sees Through” Myopia ProgressionAI doesn’t just look at “diopters” but more at ocular biometric parameters:· Monitoring Axial Length: Myopia progression is essentially axial elongation.
AI can identify previously overlooked key predictors. For example, pupil size (importance score 100) and corneal thickness are more critical than age for predicting axial growth rate .·
Extracting Image Biomarkers:
AI extracts subtle features from fundus photographs and OCT (optical coherence tomography) images that are difficult for the human eye to notice, automatically discovering “image biomarkers” related to progression .🔬
The Core Technology of Bionic EyesBionic eyes (visual prostheses) essentially bypass damaged photoreceptors and directly stimulate remaining neurons with electricity. AI plays the role of “translator,” responsible for converting camera images into electrical signals the brain can understand.·
Image Encoding and “Phosphenes”: The camera captures images, and AI processes them into stimulation patterns for phosphenes (light spots).
AI selectively highlights key information like obstacles, ground, and faces to help users navigate independently.·
Hardware Architecture Mimicking the Retina: Advanced research uses Spiking Neural Networks (SNN) to simulate the hierarchical processing of the retina (photoreceptors → bipolar cells → ganglion cells), making electrical stimulation patterns closer to natural neural signals .·
Closed-Loop Optimization: AI algorithms handle image processing while optimizing stimulation strategies in real-time based on user feedback, helping users recognize objects and navigate more easily.
Example like If you’re interested in a specific technology (such as what the images seen through bionic eyes actually look like), we can continue exploring.