Table 4 Summarized published deep learning models developed for differentiation of HPV-related dysplastic lesions in the cervical and anal regions.
Publication | AI model | Centers | Devices | Frames | Included stainings | Main characteristics of the study | Sen % | Spe % | AUC-ROC | |
|---|---|---|---|---|---|---|---|---|---|---|
Cervical lesions | Miyagi et al.16 | ResNet | 1 | 1 | 330 | Non-stained | HSIL vs LSIL only 1 frame per procedure | 80 | 88 | 0.83 |
Yuan et al.18 | ResNet | 1 | 1 | 22,330 | Non-stained acetic acid lugol | HSIL vs others (normal or LSIL) | 85 | 82 | 0.93 | |
Xue et al.17 | U-net+YOLO | 6 | 1 | 19,435 | Non-stained | HSIL vs others (normal or LSIL or cancer) | 66 | 90 | 0.78 | |
Chen et al.14 | E-B0 with GRU | 1 | 1 | 6002 | Non-stained acetic acid lugol | HSIL vs LSIL | 88 | 94 | 0.91 | |
Fang et al.15 | ShuffleNet | 1 | 1 | 1189 | Non-stained | HSIL vs others (normal or LSIL or cancer) | 82 | – | 0.99 | |
Saraiva et al.13 | ResNet | 1 | 1 | 22,693 | Non-stained acetic acid lugol | HSIL vs LSIL | 90 | 98 | 0.98 | |
Anal lesions | Saraiva et al.19 | Xception | 1 | 1 (proctoscope) | 5026 | Non-stained acetic acid lugol | HSIL vs LSIL Pilot study | 91 | 90 | 0.97 |
Saraiva et al.20 | ResNet | 1 | 1 (proctoscope) | 27,770 | Non-stained acetic acid lugol | HSIL vs LSIL Sub analysis per category, including “post-manipulation phase” (after lesion biopsy or treatment) | 97 | 94 | 0.99 | |
Saraiva et al.12 | ResNet | 2 | 2 (proctoscope+colposcope) | 57,882 | Non-stained acetic acid lugol | HSIL vs LSIL@Contributed to interoperability | 94 | 96 | 0.97 |
