Potential applications of artificial intelligence in colorectal polyps and cancer: Recent advances and prospects

Table 1 Characteristics of studies on artificial intelligence in the detection and classification of colorectal polyps

Ref.	Study type	Algorithm	Imaging modality	Image type	Training set	Testing set	Processing time
Mori et al[13]	Pilot study	-	EC	Real-time	-	-	0.3 s/image
Misawa et al[14]	Ex vivo	Machine learning: SVM	EC, NBI	Still	979 images (381 non-neoplasms, 598 neoplasms)	100 images (50 non-neoplasms, 50 neoplasms)	0.3 s/image
Kominami et al[15]	-	Machine learning: SVM	Colonoscopy, NBI	Real-time	2247 cutout training images from 1262 colorectal lesions	118 images	20 frame/s
Mori et al[16]	International web-based trial	Machine learning: SVM	EC	Still	6051 endocytoscopic images	205 small polyps (147 neoplastic and 58 non-neoplastic)	0.2 s/image
Misawa et al[17]	Pilot study	Machine learning: SVM	EC, NBI	Still	1661 EC-NBI images (1213 neoplasm images, 448 non-neoplastic images)	124 (19 neoplastic and 105 non-neoplastic)	-
Chen et al[18]	Pilot study	Deep neural network	Colonoscopy, magnifying NBI	Still	2157 (1476 neoplastic polyps vs 681 hyperplastic polyps)	284 (96 hyperplastic and 188 neoplastic polyps)	0.45 s/image
Misawa et al[19]	Ex vivo	Machine learning	Colonoscopy, WL	Video	411 (105 positive and 306 negative)	135 (50 positive and 85 negative)	-
Shin et al[20]	Pilot study	Machine learning	Colonoscopy, WL	Video	1525 (561 polyp patches and 964 normal patches)	366 (196 polyp patches and 170 normal patches)	95 ms/frame
Wang et al[21]	Ex vivo	Deep learning	Colonoscopy, WL	Still	5545 (3634 images contained polyps and 1911 images did not contain polyps)	27 113 (5541 images contained polyps and 21572 images did not contain polyps)	-
Kudo et al[22]	Pilot study	Texture analysis	EC stained or NBI image	Still	69 142 EC images (43197 stained images and 25945 NBI images)	100 polyps	0.4 s/image
Min et al[23]	Pilot study	Gaussian mixture model	Colonoscopy, linked color imaging	Still	139 images of adenomatous polyps and 69 images of non-adenomatous polyps	115 images of adenomatous polyps and 66 images of non-adenomatous polyps	-
Sánchez-Montes et al[24]	Pilot study	SVM	Colonoscopy, WL	Still	-	-	-
Horiuchi et al[25]	Pilot study	-	Colonoscopy, autofluorescence imaging	Real-time	-	-	-
Byrne et al[26]	Ex vivo	Convolutional neural network	EC, NBI	Video	223 polyp videos	125 polyp videos	50 ms/frame

EC: Endocytoscopic images; SVM: Support vector machine; NBI: Narrow-band imaging; WL: White light.

Table 2 Performance of artificial intelligence in the detection and classification of colorectal polyps

Ref.	Patients, n	Samples, n	Sensitivity, %	Specificity, %	Accuracy, %	NPV, %	PPV, %
Mori et al[13]	152	176	92.0	79.5	89.2	-
Misawa et al[14]	-	100	84.5	97.6	90.0	82.0	98.0
Kominami et al[15]	41	118	95.9	93.3	94.9	93.3	95.9
Mori et al[16]	123	205	89.0	88.0	89.0	76.0	95.0
Misawa et al[17]	58	64	94.3	71.4	87.8	83.3	89.2
Chen et al[18]	193	284	96.3	78.1	90.1	91.5	89.6
Misawa et al[19]	73	155	90.0	63.3	76.5	-	-
Shin et al[20]	-	366	95.9	95.9	95.9	-	96.4
Wang et al[21]	1138	27113	94.4	95.9	-	-	-
Kudo et al[22]	89	100	96.9 (stained)	100.0	98.0	94.6	100.0
			96.9 (NBI)	94.3	96.0	94.3	96.9
Min et al[23]	91	181	83.3	70.1	78.4	71.2	82.6
Sánchez-Montes et al[24]	-	225	92.3	89.2	91.1	87.1	93.6
Horiuchi et al[25]	77	258	80.0	95.3	91.5	93.4	85.2
Byrne et al[26]	-	106	98.0	83.0	94.0	97.0	90.0

NBI: Narrow-band imaging; NPV: Negative predictive value; PPV: Positive predictive value.