Table 2.
Comparison between classical nodule detection method and deep learning-aided method
| Method | Author | Sensitivity | False positive rate | AUC | Data set |
|---|---|---|---|---|---|
| Classical detection methods | |||||
| Conventional detection techniques | Giger et al. (1988) | 80% | 2.7 per image | –b | Authors’ compiled data set |
| Giger et al. (1990) | 57.6% | 3.7 per image | – | Authors’ compiled data set | |
| Neural network | Lo et al. (1993a) | – | – | 0.782 | Authors’ compiled data set |
| Torres et al. (2015) | 80% | 8 per scan | – | LIDC–IDRI, ITALUNG-CT, and ANODE09 | |
| Gupta et al. (2018) | 85.6% | 8 per scan | 0.957 | LIDC–IDRI | |
| 66.3% | 8 per scan | 0.831 | AAPM-SPIE-LUNGx | ||
| 70.4% | 8 per scan | 0.847 | ELCAP | ||
| 68.9 | 8 per scan | 0.804 | PCF | ||
| CNN | Lo et al. (1993b) | 80% | 2.6 per image | 0.88 | Authors’ compiled data set |
| Associative learning neural networks | Lo et al. (1998) | 80% | 2.5 per image | – | Authors’ compiled data set |
| MTANN | Suzuki et al. (2003) | 80.3% | 4.8 per scan | – | Authors’ compiled data set |
| Tajbakhsh and Suzuki (2017)a | 100% | 2.7 per scan | 0.8806 | Authors’ compiled data set | |
| KNN | Murphy et al. (2009) | 80.0% | 4.2 per scan | – | NELSON |
| SVM | Tan et al. (2011) | 83.8% | 4.0 per scan | – | LIDC |
| Setio et al. (2015) | 94.1%/98.3% | 1.0/4.0 per scan | – | LIDC–IDRI | |
| Bergtholdt et al. (2016) | 85.9% | 2.5 per scan | – | LIDC–IDRI | |
| Teramoto and Fujita (2018) | 83.5% | 5 per scan | – | East Nagoya Imaging Diagnosis Center | |
| Linear discriminant classifier + GentleBoost classifiers | Jacobs et al. (2014) | 80% | 1.0 per scan | – | NELSON |
| Deep learning-aided method | |||||
| 2D deep CNN | Setio et al. (2016) | 85.4%/90.1% | 1/4 per scan | – | LIDC–IDRI |
| Jiang et al. (2018) | 80.06%/94% | 4.7/15.1 per scan | – | LIDC–IDRI | |
| Tajbakhsh and Suzuki (2017)a | 100% | 22.7 per scan | 0.7755 | Authors’ compiled data set | |
| Nam et al. (2019) | 69.9%, 82.0%, 69.6%, and 75.0% | 0.34, 0.30, 0.02, and 0.25 per image | 0.885, 0.924, 0.831, and 0.880 | Authors’ compiled data set (evaluated on four external validation data sets) | |
| 3D deep CNN | Dou et al. (2017) | 92.2% | 8.0 per scan | – | LIDC |
| Ensemble method | Teramoto et al. (2016) | 90.1% | 4.9 per scan | – | Authors’ compiled data set |
| Setio et al. (2017) | 96.9%/98.2% | 1.0/4.0 per scan | – | LIDC–IDRI | |
aTajbakhsh and Suzuki compared performance of MTANN and CNN, which were listed in two entries in the table
bThe sign “–” means missing data