Table 6. Deep Consensus precision and recall on testing sets when trained using synthetic AND sets of different sizes with different levels of mislabelling noise.
R, ribosome data set (EMPIAR-10028); G, β-galactosidase data set (EMPIAR-10061); #Partic, number of true particles included in the data set; corrupt, corruption level. Each cell displays the precision and recall measured in each condition.
| #Partic | 3000 | 2000 | 1000 | 500 | ||||
|---|---|---|---|---|---|---|---|---|
| Corrupt | R | G | R | G | R | G | R | G |
| 30% | 0.923/0.942 | 0.918/0.914 | 0.920/0.923 | 0.902/0.917 | 0.866/0.926 | 0.876/0.952 | 0.800/0.888 | 0.816/0.900 |
| 40% | 0.839/0.902 | 0.840/0.897 | 0.774/0.819 | 0.835/0.890 | 0.738/0.820 | 0.762/0.595 | 0.693/0.818 | 0.726/0.586 |
| 45% | 0.705/0.762 | 0.695/0.817 | 0.662/0.698 | 0.610/0.701 | 0.602/0.777 | 0.581/0.731 | 0.625/0.709 | 0.578/0.604 |