Table 3.

Summary of studies about AI implementation in hysteroscopy. Sn, sensitivity; Sp, specificity; AUC—area under the curve; NK—not known; EH—endometrial hyperplasia; AH—atypical hyperplasia; EC—endometrial cancer; EP—endometrial polyps; SM—submucous myomas; FCM—Fuzzy C-Means.

Author, Year	Study Aim	Pati-ents n	Frames n	Pathologic Confirma-tion	AI Method	Dataset Method	Analysis Method	Catego-Ries	Performance Metrics %
									Sn	Sp	AUC
Neofytou, M.S.; 2006, USA [50]	Hysteroscopy image classification	198	418 frames –Normal: 209 –Abnormal: 209	No	Color– texture analysis methods	Frame annotation based on texture features (two different classifiers)	10-fold cross validation (and leave-one-out method)	Normal vs. abnormal	51–77	72–82	NK
Vlacho-kosta, 2013, Greece [51]	Differentiating normal vs. uterine vs. endometrial cancer	77	NK Only number of patients per category	Yes	DNN and FCM	Feature extraction related to vessel and texture structure	NK	Normal vs. patological	71–93	71–91	91
Zhang, 2021, China [52]	Differentiating benign (EH, EP, and SM) from premalignant/malignant lesions (AH and EC)	454	1851 frames: EH = 509 AH = 222, EC = 280 EP = 615 SM = 225	Yes	VGGNet	Image-based frame labeling with preprocessing and retaining of region of interest and data augmentation	Train–test validation (50 images for each category in test set)	Part 1: EH vs. AH vs. EC vs. EP vs. SM; Part 2: benign vs. premalignant/malignant	83	96	94
Takahashi, 2021, Japan [53]	Differentiating malign vs. benign or normal findings	177	411,800 frames –Malignant: 109,957 –Others: 301,843	Yes	Xception, MobileNetV2 and EfficientNetB0	Frame labeling in still images and video segments	Train–testvalidation	Malignant and others (uterine myoma, EP normal endometrium)	92	89	90