Table 1.
Summary of recent studies on AI systems for colorectal polyp detection.
| Author, Year | Country | Study Design | AI Algorithm | Type of Images | Outcomes |
|---|---|---|---|---|---|
| Fernández-Esparrach et al., 2016 [71] | Spain | Retrospective | WM-DOVA energy maps |
24 videos containing 31 colorectal polyps |
Sensitivity: 70.4% Specificity: 72.4% |
| Geetha et al., 2016 [72] | India | Ex vivo | Hand crafted |
Still images, 703 frames |
Sensitivity: 95% Specificity: 97% |
| Yu et al., 2017 [73] | China | Ex vivo | CNN | Videos, ASU-Mayo 18 colonoscopy videos |
Sensitivity: 71% PPV: 88% |
| Zhang et al., 2017 [74] | China | Ex vivo | CNN | Still images | Accuracy: 86% AUC: 1 |
| Billah et al., 2017 [75] | Bangladesh | Ex vivo | CNN | 14,000 still images |
Sensitivity: 99% Specificity: 99% Accuracy: 99% |
| Misawa et al., 2018 [76] | Japan | Ex vivo | CNN | Videos | Per-frame sensitivity: 90% Specificity: 63.3% Accuracy: 76.5% Per-polyp sensitivity: 94% False positive rate: 60% |
| Urban et al., 2018 [77] | United States | Ex vivo | CNN | Videos | Sensitivity: 90% |
| Figueiredo et al., 2019 [78] | Portugal | Retrospective | SVM binary classifiers |
42 colonoscopy videos containing 1680 frames with polyps and 1360 frames without polyps |
Sensitivity: 99.7% Specificity: 84.9% Accuracy: 91.1% |
| Klare et al., 2019 [79] | Germany | In vivo, prospective cohort |
KoloPol software |
Real-time colonoscopy |
Per-polyp sensitivity: 75% ADR in CADe group vs colonoscopy group: 29% vs. 31% |
| Yamada et al., 2019 [80] | Japan | Ex vivo | CNN | Videos | Sensitivity: 97.3% Specificity: 99% AUC: 0.975 |
| Wang et al., 2019 [48] | China | Prospective, RCT | EndoScreener | Real-time colonoscopy |
ADR in CADe group vs standard colonoscopy group: 29.1% vs. 20.3%, p < 0.001 |
| Liu et al., 2020 [81] | China | Prospective, RCT | Henan Tongyu |
Real-time colonoscopy |
ADR in CADe group vs control group: 39.2% vs. 24% |
| Su et al., 2020 [82] | China | Prospective, RCT |
Deep CNNs | Real-time colonoscopy |
ADR in CADe group vs control group: 28.9% vs. 16.5% |
| Ozawa et al., 2020 [83] | Japan | Ex vivo | CNN | 7077 images | Sensitivity: 92% Accuracy: 83% PPV: 86% |
| Gong et al., 2020 [84] | China | Prospective, RCT |
ENDOANGEL | Real-time colonoscopy |
ADR in CADe group vs. control group: 16% vs. 8% |
| Wang et al., 2020 [85] | China | Double-blind, RCT |
EndoScreener | Real-time colonoscopy |
ADR in CADe group (484 patients) vs control group (478 patients): 34.1% vs. 28% |
| Hassan et al., 2020 [86] | Italy | Retrospective | GI Genius | 338 videos | Per-lesion sensitivity: 99.7% |
| Repici et al., 2020 [87] | Italy | RCT | GI Genius | Real-time colonoscopy |
ADR in CADe group vs. control group: 54.8% vs. 40.4% |
AI: artificial intelligence; WM-DOVA: Window Median Depth of Valleys Accumulation; CNN: convolutional neural network; PPV: positive predictive value; AUC: area under the curve; SVM: support vector machine; ADR: adenoma detection rate; CADe: computer-aided detection; RCT: randomized controlled trial.