Comparison among CNN models
| CNN model | Train/test set | mDice | FPS (Weba) | FPS (Kerasb) | FLOPS | Parameters, n |
|---|---|---|---|---|---|---|
| mini-UNet | Human/mouse eyes | 84.0% | 23.2 | 45.2 | 0.2G | 0.03 M |
| DeepLabv3+/ResNet-50 | ImageNet + finetune on human/mouse eyes | 80.1% | <1 | 28.7 | 14.1G | 26.8 M |
| DeepLabv3+/Lite-MobileNet- V3-Small | ImageNet + fine-tune on human/mouse eyes | 69.0% | 18.8 | 34.8 | 0.3G | 1.1 M |
aDell Laptop: Intel(R) Core(TM) i7-9750H CPU @ 2.60 GHz; Intel UHD graphics 630 GPU; TensorFlow.js; backend: WebGL Browser: Microsoft edge 90.0.818.56.
bUbuntu PC: Intel(R) Core(TM) i9-9900K CPU @ 3.60 GHz; GeForce RTX 2080 Ti GPU; Python 3.6.9 + TensorFlow 2.4.1. FLOPS, FLoating point Operations Per Second; G, gigaFLOPS; M, mefaFLOPS.