Researchers propose AI breakthrough for glaucoma detection
New metric learning approach detects glaucoma progression with 0.5 dB error—more accurate than deep learning
A team of researchers from Carnegie Mellon University, University of Pittsburgh, and other institutions has developed a novel metric learning approach to improve the analysis of 3D retinal images for detecting glaucoma progression. Led by Andrew R. Cohen, the study titled *Algorithmic statistics of retinal images* introduces a method combining normalized compression distance (NCD) with anisotropic structure-enhancing filters to quantify structural differences in OCT (optical coherence tomography) scans.
The approach was validated against physician-measured visual field changes, achieving a prediction error of approximately 0.5 dB—significantly outperforming non-metric deep learning methods. The researchers also proposed normalized compression vectors (NCV) as a feature set to measure visual differences in 3D microscopy images. Their method demonstrated utility in tracking disease progression in both human patients with moderate non-progressing glaucoma and non-human primate models with manipulated intraocular pressure. The study concludes with a critique of non-metric embedding features from neural networks, highlighting their susceptibility to class-correlated statistical distortion.
- New metric learning approach using Normalized Compression Distance (NCD) outperforms deep learning in glaucoma progression detection with 0.5 dB prediction error
- Combines anisotropic structure-enhancing filters with NCD to improve accuracy in analyzing 3D retinal OCT images
- Validated on both human and non-human primate datasets, demonstrating superior performance over non-metric neural network approaches
Why It Matters
This AI-driven method could enable earlier and more accurate glaucoma diagnosis, potentially preventing vision loss through proactive intervention.