Table 2.
Prediction performance of the deep learning-based methods in terms of mean absolute error for each of the motility values and overall mean.
| Deep Learning Results | ||||
|---|---|---|---|---|
| Method | Progressive | Non-progressive | Immotile | Average Mean Absolute Error |
| Raw Frame Data Approach | ||||
| Single Frames (ResNet50) | 13.162 | 8.024 | 10.967 | 10.718 |
| Single Frames (ResNet50) + PD | 13.659 | 8.196 | 12.293 | 11.383 |
| Channel-wise Greyscale | 10.498 | 7.037 | 8.822 | 8.786 |
| Channel-wise Greyscale + PD | 11.599 | 7.849 | 10.132 | 9.860 |
| Vertical Frame Matrix | 11.149 | 8.218 | 9.418 | 9.595 |
| Vertical Frame Matrix + PD | 11.182 | 8.199 | 9.274 | 9.552 |
| Optical Flow Approach | ||||
| Sparse Optical Flow | 11.573 | 7.263 | 10.155 | 9.664 |
| Sparse Optical Flow + PD | 12.214 | 7.760 | 10.802 | 10.259 |
| Dense Optical Flow (stride = 1) | 10.191 | 7.114 | 8.914 | 8.740 |
| Dense Optical Flow (stride = 1) + PD | 10.795 | 7.856 | 8.745 | 9.132 |
| Dense Optical Flow (stride = 10) | 10.319 | 7.546 | 8.782 | 8.882 |
| Dense Optical Flow (stride = 10) + PD | 11.386 | 7.825 | 9.734 | 9.648 |
| Two Stream Network Approach | ||||
| Two Stream Sparse | 15.888 | 8.187 | 13.326 | 12.467 |
| Two Stream Sparse + PD | 16.435 | 8.197 | 13.172 | 12.601 |
| Two Stream Dense (stride = 1) | 14.583 | 7.393 | 11.996 | 11.324 |
| Two Stream Dense (stride = 1) + PD | 18.166 | 8.570 | 15.983 | 13.940 |
| Two Stream SP + DE (stride = 1) | 11.848 | 7.070 | 10.823 | 9.917 |
| Two Stream SP + DE (stride = 1) + PD | 17.304 | 8.066 | 13.783 | 13.051 |
Note that for each method, we trained two models, one with participant data and one without. Methods which used participant data under training are marked with (+PD). For the methods which use dense optical flow, stride represents the number of frames skipped when comparing the difference of two frames.