Table 1 The advancements in computer algorithms applying to cervical images
Reference | Publish year | Aim of the study | Study design | Number of subjects | Image-generating devices | Type of algorithms | Outcomes |
|---|---|---|---|---|---|---|---|
Simoes et al. [18] | 2014 | Classification of colposcopy images | Retrospective | 170 images (training set 48; test and internal validation set 122) | Digital colposcopy | ANN | Accuracy 72.15% |
Kim and Huang [19] | 2013 | Detection of CIN2+ from normal/CIN1 | Retrospective | 2000 images (normal/CIN2 1000; CIN2+ 1000) | Cervicography (discontinued) | SVM | Sensitivity 75% Specificity 75% |
Asiedu et al. [20] | 2019 | Detection of CIN1+ against normal | Retrospective | 134 patients (training set 107; internal validation set 27) | Digital colposcopy | SVM | Accuracy 80% Sensitivity 81.3% Specificity 78.6% |
Miyagi et al. [21] | 2019 | Classification of CIN1 and CIN2+ | Retrospective | 310 images (both using in training and internal validation set) | Traditional colposcopy | Convolutional neural networks | Accuracy 82.3% Sensitivity 80% Specificity 88.2% |
Song et al. [22] | 2015 | Detection of CIN2+ | Retrospective | 7669 patients with < CIN2, 142 patients with CIN2+ (training set 7531; internal validation set 280) | Cervicography (discontinued) | Multimodal convolutional neural networks | Accuracy 89% Sensitivity 83.21% Specificity 94.79% |
Schiffman et al. [23] | 2019 | Detection of CIN2+ | Retrospective | 9127 patients with < CIN2, 279 patients with CIN2+ (training set 744, internal validation set 324, rest in screening set) | Cervicography (discontinued) | Faster R-CNN | AUC 0.91 |