Table 6.
Performances of the artificial intelligence studies.
| Therapy | AI | Signature | Training Set | Test Set | Reference |
|---|---|---|---|---|---|
| Chemotherapy | ML | Delta-radiomics | Excellent, 99% correctly classified as responders, 94% correctly classified as non-responders | Excellent, 85% correctly classified as responders, 92% correctly classified as non-responders | [75] |
| ML | Radiomics | Good | Poor | [79] | |
| DL | Fusion radiomics | Excellent, 84.7% correctly classified as responders, 84.8% correctly classified as non-responders (radiomics model) Excellent, 89.8% correctly classified as responders, 84.8% correctly classified as non-responders (fusion radiomics model) |
Poor, 90% correctly classified as responders, 73% correctly classified as non-responders (radiomics model) Good, 90.9% correctly classified as responders, 73.3% correctly classified as non-responders (fusion radiomics model) |
[80] | |
| ML | Radiomics | 76% correctly classified as responders, 67% correctly classified as non-responders (per-lesion 3D set) | 61% correctly classified as responders, 60% correctly classified as non-responders (per-lesion 3D set) 41% correctly classified as responders, 21% correctly classified as non-responders (per-patient 3D set) |
[81] | |
| ML | Radiomics | Poor (high standard quality computed tomography scan, HQCT, set) Poor (high standard quality computed tomography scan, SQCT, set) |
Failed (HQCT set) Failed (SQCT set) |
[87] | |
| ML | 74 genes | 97.6% correctly classified as responders, 100% correctly classified as non-responders | [95] | ||
| ML | 16 long noncoding RNA | Good | [94] | ||
| ML | 16 long noncoding RNA | Poor | [94] | ||
| ML | 16 long noncoding RNA | Poor | [94] | ||
| Targeted therapy | DL | Radiomics | Good, 91.7% correctly classified as responders, 75% correctly classified as non-responders (four features model) Good, 91.9% correctly classified as responders, 75% correctly classified as non-responders (four features model) |
[98] | |
| DL | Radiomics | Good, 97% correctly classified as responders, 59% correctly classified as non-responders | Good, 98% correctly classified as responders, 54% correctly classified as non-responders | [86] | |
| ML | Radiomics | Good, 77% correctly classified as responders, 85% correctly classified as non-responders (high quality, HQ, cohort) Good (standard quality, SQ, cohort) |
Good, 80% correctly classified as responders, 78% correctly classified as non-responders (HQ cohort) Poor, 82% correctly classified as responders, 61% correctly classified as non-responders (SQ cohort) |
[87] | |
| ML | Radiomics | 89% correctly classified as responders, 85% correctly classified as non-responders (lesion model) | 90% correctly classified as responders, 42% correctly classified as non-responders (lesion model) 92% correctly classified as responders, 86% correctly classified as non-responders (patient model) |
[74] | |
| ML | 16-long noncoding RNA | Poor | [94] | ||
| ML | 16-long noncoding RNA | Worthless | [94] | ||
| ML | 16-long noncoding RNA | Poor | [94] | ||
| ML | 67-gene | Good | Poor | [99] | |
| Chemotherapy and targeted therapy | ML/DL | Imaging/molecular signatures | Good | Good | Present meta-analysis |
ML, machine learning; and DL, deep learning.