LightVesselNet: 75K-parameter AI detects eye disease on edge devices
A sub-100K parameter model matches giants in retinal vessel segmentation.
LightVesselNet is an ultra-lightweight neural network for retinal blood vessel segmentation, a critical step in early detection of diabetic retinopathy and glaucoma. With just 75,000 parameters—far fewer than typical deep learning models that require heavy computational resources—it uses a compact encoder-decoder architecture enhanced with channel and spatial attention mechanisms. A multi-scale feature aggregation module at the bottleneck captures vessels of varying widths, while a subpixel upsampling strategy in the decoder preserves resolution. A dedicated edge residual connection ensures fine vessel detail is maintained throughout decoding.
The model was tested on five public datasets: DRIVE, STARE, CHASEDB1, FIVES, and HRF. Results show sensitivity scores ranging from 0.8096 to 0.8640 and Dice coefficients from 0.7686 to 0.8649, rivaling much larger state-of-the-art models. Cross-dataset evaluation confirmed strong generalization. LightVesselNet's efficiency (performance relative to parameters and FLOPs) makes it a strong candidate for deployment in low-resource clinical settings and mobile screening tools, where real-time, accurate diagnostics are needed without cloud dependency.
- LightVesselNet achieves competitive retinal vessel segmentation with only 75K parameters.
- Tested on five datasets (DRIVE, STARE, CHASEDB1, FIVES, HRF) with sensitivity up to 0.8640 and Dice up to 0.8649.
- Compact architecture uses attention, multi-scale features, subpixel upsampling, and edge residual connections.
Why It Matters
Enables accurate eye disease screening on low-cost mobile devices, expanding access in underserved clinics.