Table 3.
Presentation of our method and several previous studies in terms of methods, datasets, and resulting classification AUCs
| Method | Training #Patients | Test #Patients | #EGFR muta | AUC (%) |
|---|---|---|---|---|
| Radiomics23 | 353 | 352 | 183 (24.0%)b | 69 |
| + clinical information | 75 | |||
| Radiomics17 | 298 | NAc | 137 (46.0%) | 64.7 |
| + clinical information | 70.9 | |||
| Radiomics18 | 47 | NAc | 19 (40.4%) | 67.0 |
| Radiomics (This study) | 464 (HdH train‐dev Dataset) | 115 (HdH test Dataset) | 62 (53.9%) | 64.5 |
| 3D DenseNets w/mixup, ensemble (This study) | 348 (HdH training Dataset) | 115 (HdH test Dataset) | 62 (53.9%) | 75.8 |
| Radiomics (this study) | 464 (HdH train‐dev Dataset) | 37 (TCIA Dataset) | 9 (24.3%) | 68.7 |
| 3D DenseNets w/mixup, ensemble (this study) | 348 (HdH training Dataset) | 37 (TCIA Dataset) | 9 (24.3%) | 75.0 |
a
Shown as the number of cases (percentage).
b
Estimated using the proportion of EGFR Mut on the entire data set, rather than the test set.
c
The evaluation results are based on multivariate statistical analysis, rather than the practice of training – validation (development) – test in machine learning. Since the prior studies listed in the above table used nonshared datasets independently, the results are for reference only.