Table 1.
Deep learning for computer-aided gastrointestinal endoscopy: target disease, method, dataset and outcome summaries of selected comprehensive studies.
| Type proc. | Organ | Mod. | Target disease | Dataset | Method | Outcome | Similar studies |
|---|---|---|---|---|---|---|---|
| OGD | O | WL | BE |
Train: 494,364 Test: 1704 (669 patients) |
Classificationa1—Neoplasia vs NDBE (hybrid ResNet-UNet) |
(DS 4) sensitivity: 90%, specificity: 88%, accuracy: 89% (DS 5) sensitivity: 93%, specificity: 83%, accuracy: 88% |
Ebigbo et al.2 (ResNet100) |
| OGD | O | NBI | SCC |
Train: 6473 images Test: 6671 images and 80 videos |
Segmentation39 (SegNet) |
(Per-image) sensitivity: 98.04%, specificity: 95.03% (Per-frame) sensitivity: 91.5%, specificity: 99.9% |
Nakagawa et al.116, Sho et al.117 (SSD) Everson et al.5 (Deep supervision) |
| OGD | S | WLI | AG |
5470 images Train: 70% Test: 30% |
Classification3 (DenseNet121) | Sensitivity: 94.5%, specificity: 94%, accuracy: 94.2% | Guimarães et al.4 (VGG16) |
| OGD | S | WLI | AG, IM, erosion and hem. |
Train: 7326 images Val: 815 images Test: 570 images, 258 external test and 80 videos |
Classificationa41 (UNet++, ResNet50) | Accuracy (non AG/AG, atrophy/IM, and erosion/haemorrhage): 88.78%, 87.40% and 93.67% (int. test), 91.23%, 85.81% and 92.70% (ext. test) and 95.00 %, 92.86 %, and 94.74% (video) | Zhao et al.94 (UNet)b,c |
| Colon | CR | WL | Polyp |
Train: 411 clips Test: 135 clips (videos) |
Frame-level polyp/non-polyp classification42 (3D CNN, binary) | Sensitivity: 90%, specificity: 63%, accuracy: 76%, FP: 60 | Kim et al.118 (TL: AlexNet) |
| Colon | CR | WL, NBI | Polyp |
Train: 8641 images Test: 1330 images and 11 videos |
Polyp detection with localisation43 (YOLO; VGG16 (A1), VGG19 (A2) and ResNet50 (A3)) | (A2) sensitivity: 90%, specificity: 95.2%, AUC: 0.991, accuracy: 96%, FP: 7 |
Yamada et al.119 (Faster R-CNN) Klare et al.c95 |
| Colon | CR | WL, NBI | Polyp |
Train: 20,431 images Test: 7077 images (1172 polyps) |
Detection6 for polyp characterisation (SSD) |
(WL) sensitivity: 90%, PPV: 83% (NBI) sensitivity: 97%, PPV: 98% |
Lee et al.120d Zachariah et al.121 |
| Colon | CR | NBI | Polyp |
Train: 1100 (adem.) and 1050 (hyp.) Test: 300 images (180: adem. and 120 hyp.) |
Classification9 for polyp characterisation (AutoML) | Sensitivity: 83.3%, specificity: 91.7%, accuracy: 86.7% | Song et al.8 (CNN) Byrne et al.7 (CNN) |
| Colon | CR | WL | IBD (UC) |
1651 images Train: 80% Val: 10% Test: 10% and 30 videos |
Classification11 into MCES scoring (159-layer CNN) |
Sensitivity: 83%, specificity: 96% PPV: 86%, NPV: 94% |
Ozawa et al.44 (GoogLeNet) Becker et al.a45 (CNN) |
| Colon | CR | WL | CRC |
Train: 464,105 Test: TCH: 20,783, TFCH: 15,441 and TGH: 48,391 |
Classification48 benign/malignant (169-layer DenseNet) (CRCNet) |
(Test set: sensitivity, specificity) TCH: 90.4%, 85.3% TFCH: 78.9%, 95.0% TGH: 74.6%, 99.2% |
Ito et al.122 (AlexNet) |
OGD oesophago-gastro-duodenoscopy, DNN deep neural network, CNN convolutional neural network, WLI white light imaging, NBI narrow band imaging, PPV positive predictive value, NPV negative predictive value, O oesophagus, CR colorectal, IBD inflammatory bowel disease, UC ulcerative colitis, MCES Mayo Clinic Endoscopic Subscore, SSD Single Shot MultiBox Detector, A1–A3 architectures from 1 to 3, TCH Tianjin Cancer Hospital, TFCH Tianjin First Central Hospital, TGH Tianjin General Hospital.
aMultisite study.
bComparative: DL vs endoscopists.
cProspective study.
dPublic dataset.