TABLE 2.
Summary of AI applications in TB detection.
| No. | References | Method | Reference standard | Dataset | Study population | Training/Validation/test cohort | Model names | Algorithm | Results |
| 1 | Lakhani and Sundaram (31) | Retrospective multi-center on CXR images | Sputum, radiology reports, radiologists, and clinical records. | 1,007 participants | United States, China, and Belarus | Training: 685 Validation: 172 Test: 150 | NA | CNN | AUC 0.99, Sen 97.3%, Spe 94.7%, Acc 96.0% of the ensemble method |
| 2 | Hwang et al. (32) | Retrospective multi-center on CXR images | Culture or PCR | 62,433 CXR images | Korea, China, United States, etc. | Training: 60,089 Tuning: 450 Internal validation: 450 External validation: 1,444 | DLAD | CNN | AUC 0.977–1.000 for TB classification, AUAFROC 0.973–1.000 for lesion localization; Sen 0.943–1.000, Spe 0.911–1.000 at high sensitivity cutoff |
| 3 | Nijiati et al. (33) | Retrospective single-center on CXR images | Symptoms, laboratory and radiological examinations | 9,628 CXR images | China | Training: 7,703 Test: 1,925 | NA | CNN | AUC 0.9902–0.9944, Sen 93.2–95.5%, Spe 95.78–98.05%, Acc 94.96–96.73% in the test set |
| 4 | Lee et al. (34) | Retrospective single-center on CXR images | Smear microscopy, culture, PCR, and radiologists | 19,686 participants | Korea | Test: 19,686 | DLAD | CNN | AUC 0.999, Sen 1.000, Spe 0.959–0.997, Acc 0.96–0.997 |
| 5 | Heo et al. (35) | Retrospective single-center on CXR images | Radiologists | 39,677 participants | Korea | Training: 2,000 Test: 37,677 | D-CNN and I-CNN | CNN | AUC 0.9213, Sen 0.815, Spe 0.962 of D-CNN |
| 6 | Nafisah and Muhammad (36) | Retrospective multi-center on CXR images | NA | 1,098 CXR images | United States, China, and Belarus | 5-fold cross validation | NA | CNN | AUC 0.999, Acc 98.7%, recall 98.3%, precision 98.3%, Spe 99.0% |
| 7 | Pasa et al. (37) | Retrospective multi-center on CXR images | NA | 1,104 participants | United States, China, and Belarus | 5-fold cross validation | NA | CNN | AUC 0.925, Acc 86.2% |
| 8 | Rajaraman et al. (38) | Retrospective multi-center on CXR images | Radiologists | 76,031 CXR images | United States and Spain | Training: test 9:1 | NA | CNN | AUC 0.9274–0.9491, recall 0.7736–0.8113, precision 0.9524–0.9773, Acc 0.8585–0.8962 |
| 9 | Rajpurkar et al. (39) | Retrospective multi-center on CXR images | Culture or Xpert MTB/RIF | 677 participants | South Africa | Training: 563 Test: 114 | CheXaid | Deep learning | AUC 0.83, Sen 0.67, Spe 0.87, Acc 0.78 |
| 10 | Lee et al. (40) | Retrospective multi-center on CXR images | Sputum microscopy, culture or PCR | 6,964 participants | Korea | Training: validation 7:3 Test: 455 | NA | CNN | AUC 0.82–0.84, Spe 26–48.5% at the cutoff of 95% Sen in the test set |
| 11 | Yan et al. (41) | Retrospective multi-center on CT images | Culture | 1,248 CT images | China and United States | Training: validation 8:2 External test: 356 | NA | CNN | Acc 95.35–98.25%, recall 94.87–100%, precision 94.87–98.70% |
| 12 | Khan et al. (43) | Prospective single-center on CXR images | Culture | 2,198 participants | Pakistan | Test: 2,198 | qXR and CAD4TB | CNN | AUC 0.92, Sen 0.93, Spe 0.75 for qXR; AUC 0.87, Sen 0.93, Spe 0.69 for CAD4TB |
| 13 | Qin et al. (44) | Retrospective multi-center on CXR images | Xpert MTB/RIF | 1,196 participants | Nepal and Cameroon | Test: 1,196 | qXR, CAD4TB, and Lunit INSIGHT CXR | CNN | AUC 0.92–0.94, Sen 0.87–0.91, Spe 0.84–0.89, Acc 0.85–0.89 |
| 14 | Qin et al. (45) | Retrospective multi-center on CXR images | Xpert MTB/RIF | 23,954 participants | Bangladesh | Test: 23,954 | qXR, CAD4TB, InferRead DR, etc. | CNN | AUC 84.89–90.81%, Sen 90.0–90.3%, Spe 61.1–74.3% when fixed at 90% Sen |
| 15 | Codlin et al. (46) | Retrospective multi-center on CXR images | Xpert MTB/RIF | 1,032 participants | Viet Nam | Test: 1,032 | qXR, CAD4TB, Genki, etc. | CNN | AUC 0.50–0.82, Spe 6.3–48.7%, Acc 17.8–54.7% when fixed at 95.5% Sen |
| 16 | Melendez et al. (47) | Retrospective single-center on CXR images | Culture | 392 patients | South Africa | 10-fold cross validation | CAD4TB | Machine learning | AUC 0.72–0.84, Spe 24–49%, NPV 95–98% when fixed at 95% Sen |
AI, artificial intelligence; TB, tuberculosis; CXR, chest X-ray; NA, not available; CNN, convolutional neural network; AUC, area under the curve; Sen, sensitivity; Spe, specificity; Acc, accuracy; PCR, polymerase chain reaction; AUAFROC, area under the alternative free-response receiver-operating characteristic curve; CT, computed tomography.