1 Introduction
-
A C. elegans skeletonization method is proposed based on U-Net type neural networks with low-resolution images and noise.
-
A new method for generating low-resolution synthetic images is proposed to easily generate a custom-labeled dataset for different C. elegans behaviors.
-
A neural network has been trained with a low-resolution synthetic image and successfully tested in the domain of real images.
-
Different U-Net architectures have been compared with an algorithm based on traditional image processing techniques.
2 Related Work
2.1 Caenorhabditis elegans and Neural Networks
2.2 U-Nets
2.3 Data Augmentation and Synthetic Images
3 Methods
3.1 Strain and Culture Conditions of C. elegans
3.2 Real-Image Acquisition Method
3.3 Image Simulation Method
3.4 Classical Skeletonization Method
3.5 Skeletonization Method Using Improved Skeleton
3.6 Proposed Skeletonization Method
4 Evaluation Method
5 Experiments and Results
5.1 Method Comparison
Avg. worm-ends class | Avg. worm-body class | Total average worm class | |||||
---|---|---|---|---|---|---|---|
Model | loss | IoU | E.D | IoU | E.D | IoU ± IC 95% | E.D. ± IC 95% |
U-Net | 6.47E-06 | 0.8883 | 0.1875 | 0.9315 | 0.0747 | 0.9322±2.40E-04 | 0.1067±5.12E-04 |
U-Net A | 3.05E-04 | 0.8179 | 0.1262 | 0.8573 | 0.1269 | 0.8597±1.24E-03 | 0.1270±1.06E-03 |
UMF U-Net | 1.54E-06 | 0.9317 | 0.0019 | 0.9371 | 0.0006 | 0.9361±9.30E-04 | 0.0009±6.99E-05 |
SmaAt DS | 2.16E-05 | 0.9153 | 0.0241 | 0.9297 | 0.0184 | 0.9287±1.09E-03 | 0.0199±3.68E-04 |
SmaAt DS AT | 1.60E-05 | 0.9191 | 0.0226 | 0.9305 | 0.0112 | 0.9295±1.03E-03 | 0.0141±3.38E-04 |
SmaAt DS AT 4C | 1.23E-04 | 0.8120 | 0.1263 | 0.8625 | 0.1073 | 0.8615±1.07E-03 | 0.1123±7.15E-04 |
SmaAt AT | 3.06E-06 | 0.9324 | 0.0017 | 0.9364 | 0.0011 | 0.9355±9.42E-03 | 0.0012±9.44E-05 |
Avg. aggregation | Avg. Agg. with noise | Avg. rolled | Total average worm | ||||||
---|---|---|---|---|---|---|---|---|---|
Model | N parameters | IoU | E.D | IoU | E.D | IoU | E.D | IoU ± CI 95% | E.D. ± CI 95% |
ISA | 0.7625 | 0.5659 | 0.6421 | 2.1726 | 0.8122 | 0.5540 | 0.6936±0.0125 | 1.6185±0.1517 | |
U-Net | 17.2576M | 0.7634 | 0.5551 | 0.6510 | 0.9001 | 0.7600 | 0.6678 | 0.6923±0.0134 | 0.8092±0.0466 |
U-Net A | 17.2664M | 0.6858 | 0.7822 | 0.6980 | 0.6712 | 0.7172 | 0.6089 | 0.6977±0.0105 | 0.6274±0.0277 |
UMF U-Net | 17.2664M | 0.6992 | 0.6090 | 0.7313 | 0.7259 | 0.7622 | 0.5686 | 0.7279±0.0066 | 0.6097±0.0276 |
SmaAt DS | 3.9536M | 0.6339 | 0.5737 | 0.7133 | 0.6562 | 0.7481 | 0.5905 | 0.6993±0.0086 | 0.6529±0.0323 |
SmaAt DS AT | 4.0320M | 0.5759 | 0.5870 | 0.6849 | 0.6505 | 0.7004 | 0.6150 | 0.6613±0.0107 | 0.6642±0.0347 |
SmaAt DS AT 4C | 3.9986M | 0.5713 | 0.7471 | 0.7193 | 0.7032 | 0.7497 | 0.6053 | 0.6884±0.0125 | 0.6240±0.0194 |
SmaAt AT | 17.3447M | 0.6767 | 0.8359 | 0.7338 | 0.6459 | 0.7610 | 0.6146 | 0.7240±0.0082 | 0.6476±0.0232 |
5.2 Real Versus Synthetic Image Training
Train | Test | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
N | Annotated | Best | Avg. aggregation | Avg. Agg. with noise | Avg. rolled | Total average | |||||
worms | poses | epoch | IoU | E. D | IoU | E. D | IoU | E. D | IoU ± CI 95% | E. D. ± CI 95% | |
Real | 627 | 18810 | 11 | 0.4987 | 1.7913 | 0.3965 | 3.5523 | 0.4405 | 2.3572 | 0.4265±0.0094 | 2.9776±0.1403 |
Synthetic | 1 | 30 | 14 | 0.6358 | 0.6487 | 0.6914 | 0.7795 | 0.7424 | 0.6146 | 0.6850±0.0115 | 0.7267±0.0419 |
Synthetic | 16 | 480 | 13 | 0.6782 | 0.6581 | 0.7097 | 0.7346 | 0.7264 | 0.6554 | 0.7044±0.0077 | 0.6824±0.0323 |
Synthetic | 30 | 900 | 9 | 0.6997 | 0.5777 | 0.7267 | 0.7299 | 0.7432 | 0.6387 | 0.7225±0.0082 | 0.6818±0.0332 |
Synthetic | 627 | 18810 | 5 | 0.6992 | 0.6090 | 0.7313 | 0.7259 | 0.7622 | 0.5686 | 0.7279±0.0066 | 0.6097±0.0276 |