Table 3.
Summary estimate of pooled performance of artificial intelligence in image-based lung cancer diagnosis
| Subgroup | No. of studies (datasets) | Tau2 | Sensitivity | Specificity | P valueb | ||||
|---|---|---|---|---|---|---|---|---|---|
| Sensitivity | P valuea | I2 (95%CI) | Specificity | P valuea | I2 (95%CI) | ||||
| Overall | 209 (251) | 0.0063 | 0.86 (0.84–0.87) | < 0.05 | 94.71 (94.30–95.12) | 0.86 (0.84–0.87) | < 0.05 | 97.35 (97.19–97.51) | |
| Objective | 0.0001 | ||||||||
| Malignant/benign | 128 (151) | 0.0058 | 0.88 (0.86–0.90) | < 0.05 | 96.53 (96.23–96.84) | 0.88 (0.85–0.90) | < 0.05 | 98.37 (98.26–98.48) | |
| ADC/SCC | 19 (22) | 0.0243 | 0.81 (0.76–0.85) | < 0.05 | 77.03 (67.73–86.33) | 0.80 (0.74–0.84) | < 0.05 | 76.20 (66.47–85.93) | |
| Invasive/pre-invasive | 16 (23) | 0.0256 | 0.86 (0.82–0.89) | < 0.05 | 62.91 (46.28–79.55) | 0.82 (0.79–0.84) | < 0.05 | 40.33 (10.58–70.08) | |
| EGFR mutant/wild | 46 (55) | 0.0291 | 0.78 (0.75–0.81) | < 0.05 | 73.50 (66.50–80.51) | 0.81 (0.77–0.84) | < 0.05 | 85.49 (82.25–88.73) | |
| Cohort | 0.0231 | ||||||||
| Internal validation cohort | (195) | 0.0113 | 0.86 (0.85–0.88) | < 0.05 | 93.93 (93.38–94.48) | 0.86 (0.84–0.88) | < 0.05 | 96.54 (96.28–96.80) | |
| External validation cohort | (56) | 0.0622 | 0.82 (0.78–0.86) | < 0.05 | 96.35 (95.81–96.88) | 0.84 (0.79–0.88) | < 0.05 | 98.61 (98.47–98.76) | |
| Algorithm | 0.0630 | ||||||||
| Machine learning | 114 (136) | 0.0616 | 0.84 (0.82–0.86) | < 0.05 | 88.58 (87.07–90.08) | 0.83 (0.81–0.86) | < 0.05 | 96.35 (96.01–96.69) | |
| Deep learning | 95 (115) | 0.0054 | 0.87 (0.85–0.89) | < 0.05 | 96.98 (96.69–97.27) | 0.87 (0.85–0.89) | < 0.05 | 97.41 (97.18–97.65) | |
| 3D deep learning | 19 (22) | 0.0845 | 0.87 (0.82–0.90) | < 0.05 | 97.26 (96.68–97.85) | 0.89 (0.85–0.92) | < 0.05 | 96.31 (95.43–97.18) | |
| 2D deep learning | 76 (93) | 0.0018 | 0.87 (0.85–0.90) | < 0.05 | 96.70 (96.34–97.06) | 0.87 (0.84–0.89) | < 0.05 | 97.71 (97.48–97.93) | |
| Imaging variable only | 0.7667 | ||||||||
| Yes | 133 (155) | 0.0079 | 0.87 (0.85–0.88) | < 0.05 | 96.07 (95.65–96.48) | 0.87 (0.84–0.89) | < 0.05 | 98.36 (98.25–98.47) | |
| No | 76 (96) | 0.0315 | 0.83 (0.81–0.85) | < 0.05 | 84.28 (81.57–86.99) | 0.83 (0.80–0.85) | < 0.05 | 85.52 (83.08–87.96) | |
| Nodule segmentation | 0.7535 | ||||||||
| Manual | 78 (94) | 0.0331 | 0.84 (0.82–0.86) | < 0.05 | 85.31 (82.80–87.82) | 0.82 (0.80–0.85) | < 0.05 | 89.91 (88.36–91.46) | |
| Algorithm | 131 (157) | 0.008 | 0.86 (0.84–0.88) | < 0.05 | 96.32 (96.00–96.64) | 0.87 (0.84–0.89) | < 0.05 | 98.22 (98.09–98.34) | |
| Image quality control | 0.1355 | ||||||||
| Yes | 63 (72) | 0.019 | 1.00 (0.99–1.00) | < 0.05 | 95.14 (94.46–95.83) | 0.85 (0.82–0.88) | < 0.05 | 97.93 (97.71–98.15) | |
| No | 145 (179) | 0.01 | 0.86 (0.84–0.88) | < 0.05 | 94.97 (94.52–95.43) | 0.86 (0.83–0.87) | < 0.05 | 96.54 (96.27–96.82) | |
aP-Value for heterogeneity within each subgroup.
bP-Value for heterogeneity between subgroups with multivariable meta-regression analysis.