Table 1.
Image matching performance on the openly available GRAPE dataset3
| Image-level | Patient-level | |||||
|---|---|---|---|---|---|---|
| R@1 | R@10 | AP | R@1 | R@10 | AP | |
| RETFound (From the main results table in Nebbia et al.) | 38.94% | 56.11% | 83.53% | 54.72% | 72.64% | 83.60% |
| ResNet10t (Tiny CNN that has never seen a retinal image before) | 72.04% | 83.72% | 96.74% | 82.08% | 91.51% | 98.38% |
| Raw pixels (16 × 16 resolution) | 25.13% | 36.46% | 78.36% | 42.45% | 59.43% | 70.55% |
For the raw pixel values, we simply resize the images to a resolution of 16 × 16 pixels and flatten the image into a single vector of length 768 (16 × 16 pixels * 3 colour channels). R Recall, AP Average Precision.