Tiny U-Net Paired With Classic Clustering Beats Deep Learning on Eye Lesions
A hybrid U-Net plus SLIC clustering pipeline hits 89.7% Dice on 1024x1024 eye fundus scans, quadrupling Attention U-Net accuracy on high-res lesion segmentation.
Takeaways
- Hybrid model hits 89.7% Dice / 80.1% IoU on 1024x1024 fundus images, versus 51.3% / 34.2% for Attention U-Net.
- Small 128x128 U-Net generates a coarse mask that automatically configures SLIC superpixel clustering on the full-resolution image.
- Runs inference on CPU and scales to native 3900x3900 fundus captures without additional training.
- Generalizes across cameras: 76.7% Dice on external Wills Eye Hospital data versus 54.1% for U-Net.
- Main failure mode is when the small U-Net detects nothing, leaving SLIC without a seed.
- Full paper on arXiv; code and dataset are not publicly released.
Segmenting tiny lesions in high-resolution medical images is one of those problems where deep learning quietly falls apart. A new paper on choroidal nevi (pigmented spots at the back of the eye that can, rarely, turn into melanoma) shows that a plain Attention U-Net trained on 1024x1024 fundus photos lands at a Dice coefficient of just 51.3%, while the same architecture on 128x128 versions of the same images hits 84.3%. The authors propose a hybrid pipeline that keeps the small U-Net and hands its output to a classical clustering algorithm, pushing Dice to 89.7% on the large images while running on a CPU.
Why U-Net collapses on big fundus images
Choroidal nevi lesions occupy less than 2% of the entire fundus image on average, so as resolution grows, the amount of irrelevant retinal background grows quadratically while the signal stays fixed. The team ran a controlled experiment across four resolutions and observed a clean monotonic decay:
| Image size | IoU | Dice | Train time | Energy |
|---|---|---|---|---|
| 128x128 | 74.82% | 84.32% | 3.16 min | 15.8 Wh |
| 256x256 | 66.77% | 79.70% | 11.2 min | 56 Wh |
| 512x512 | 46.44% | 62.43% | 37 min | 185 Wh |
| 1024x1024 | 34.20% | 51.30% | 129.3 min | 646.5 Wh |
Training a 1024x1024 model on a V100 burned 40x more energy than the 128x128 run and produced worse masks. The authors attribute this to three compounding effects: worsening signal-to-noise ratio, uneven noise distribution across larger frames, and the limited capacity of U-Net itself, which cannot simply be made deeper without hitting vanishing gradients.