Could artificial intelligence-powered colonoscopies change the future of colorectal cancer screening?

Table 1 Trials using computer-aided detection system for colorectal lesions[86-108]

Ref.	Year	Application used	Patients (n)	Findings
Wang et al[86]	2019	Endo Screener	1058	AI system significantly increased ADR (29.1% vs 20.3%; P < 0.001)
Su et al[87]	2019	AQCS	623	AI significantly increased ADR (28.9% vs 16.5%; P < 0.001)
Liu et al[88]	2020	Henan Tongyu	1026	AI significantly increased ADR (39% vs 24%; P < 0.001)
Gong et al[89]	2020	ENDOANGEL	704	AI significantly increased ADR (16% vs 8%; P = 0.001)
Repici et al[90]	2020	GI genius	685	AI significantly increased ADR (54.8% vs 40.4%)
Wang et al[91]	2020	Endo Screener	369	AMR was significantly lower with AI (13.9% vs 40.0%; P < 0.001)
Wang et al[92]	2020	Endo Screener	962	AI system significantly increased ADR rather than the sham system (34% vs 28%; P = 0.030)
Kamba et al[93]	2021	Endo Screener	358	AMR was significantly lower with AI (13.8% vs 40.6%; P < 0.001)
Xu et al[94]	2021	Endo Screener	2325	AI system did not significantly increase PDR (38.8% vs 36.2%; P = 0.183)
Luo et al[95]	2021	Endo Screener	150	AI system significantly increased PDR (38.7% vs 34.0%; P < 0.001)
Shaukat et al[96]	2022	Endo Screener	1359	AI system significantly increased adenomas per colonoscopy (1.05 vs 0.83; P = 0.002)
Wallace et al[97]	2022	GI genius	230	AMR was significantly lower with AI (15.5% vs 32.4%; P < 0.001)
Glissen Brown et al[98]	2022	Endo Screener	223	AMR was significantly lower with AI than with high-definition white light colonoscopy (20.12% vs 31.25%; P = 0.025)
Lui et al[99]	2023	Endo Screener	216	AI significantly increased ADR in proximal colon (44.7% vs 34.6%)
Ahmad et al[100]	2023	GI genius	614	AI system did not significantly increase ADR (71.4% vs 65.0%; P = 0.09)
Mangas-Sanjuan et al[101]	2023	GI genius	3213	AI system did not significantly increase advanced colorectal neoplasia detection rate (34.8% vs 34.6%; P = 0.91)
Karsenti et al[102]	2023	GI genius	2015	AI system slightly increased ADR (37.5% vs 33.7%; P = 0.051)
Wei et al[103]	2023	Endo Vigilant	769	AI system did not significantly increase ADR (35.9% vs 37.2%; P = 0.774)
Nakashima et al[104]	2023	CAD EYE	415	AI system significantly increased ADR (59.4% vs 47.6%; P = 0.018)
Gimeno-García et al[105]	2024	ENDO-AID	370	AI system significantly increased ADR (55.1% vs 43.8%; P = 0.029)
Yao et al[106]	2024	ENDOANGEL	685	AMR was significantly lower with AI (18.82% vs 43.69%; P < 0.001)
Schöler et al[107]	2024	CAD EYE	286	AI system did not significantly increase ADR (43% vs 41%; P = 0.696)
Yamaguchi et al[108]	2024	CAD EYE	231	AMR was significantly lower with AI (25.6% vs 38.6%; P = 0.033)

AQCS: Automatic quality control system; GI: Gastrointestinal; CAD: Computer-aided diagnosis; ENDO-AID: Endoscopy computer-aided diagnosis system; AI: Artificial intelligence; ADR: Adenoma detection rate; PDR: Polyp detection rate; AMR: Adenoma miss rate.

Table 2 Performance comparison of studied models[110]

Model (year)	CVC-ClinicDB				Kvasir-SEG				CVC-ColonDB				EndoScene
	Dice	IoU	Rec	Prec	Dice	IoU	Rec	Prec	Dice	IoU	Rec	Prec	Dice	IoU	Rec	Prec
UNet (2015)	29.83	19.57	75.07	22.76	39.76	27.53	84.78	31.28	15.38	9.87	85.321	14.53	29.21	19.54	58.73	25.88
U2Net (2020)	91.06	84.61	91.58	92.322	86.52	77.45	85.89	89.8	82.28	73.43	79.53	88.432	90.96	85.532	91.23	92.172
I2UNet (2024)	92.322	87.602	94.141	78.01	87.75	84.451	86.70	88.59	75.05	65.94	69.89	49.50	87.41	77.87	92.382	77.53
SR-AttNet (2023)	85.20	76.79	88.92	84.14	88.02	80.83	88.372	90.892	83.352	76.422	84.09	83.41	91.032	85.04	90.99	92.02
FoccusU2Net (proposed)	93.61	89.31	93.12	93.31	89.81	84.32	88.982	91.142	86.41	77.61	84.22	93.21	93.61	88.11	93.71	94.81

¹The best values.

²The second-best values.

Dice score: A metric used to measure the resemblance of two groups; IoU: A metric used to evaluate the accuracy of annotation, segmentation and object detection algorithms; Rec: Recall the true positive rate; Prec: Precision the degree of accuracy.

Table 3 The comparison between ileo-colonoscopy and colon capsule endoscopy[128-138]

	CCE	Colonoscopy
Extent of gastrointestinal tract examined	Gastric antrum, small bowel and colon on CCE	Terminal ileum and colon only for colonoscopy
Patient safety	CCE has non-invasive with minimal capsule retention risk, reliant on patient selection (0.73%-2%)	Colonoscopy is invasive with perforation risk: 88 per 100000 people (0.88%)
Bowel preparation requirement	Additional low residue diet or high-volume laxative e.g., polyethylene glycol in addition to standard bowel preparation for CCE	Standard bowel preparation including volume bowel preparation in standard colonoscopy
Ability in taking biopsies and therapy	Unable to take biopsies or perform therapeutics with capsule	Able to take biopsies or perform therapeutics with the colonoscope
Localization	No scope guided for localization of pathology other than visual landmarks such as ileocecal valve, appendiceal orifice and anal cushion in video-capsule	Scope guided is available for more accurate localization of the pathologies within the colon at colonoscopy
Procedure time	CCE has an average reading time: 45-60 minutes	Colonoscopy has an average 30 minutes procedural slots

CCE: Colon capsule endoscopy.

Table 4 Real-life recommendations of world-wide major associations[148-153]

Organization	Starting age	Screening methods	Interval	AI integration/position
American Cancer Society	45	Colonoscopy, stool-based (FIT/FIT-DNA), CT-colonography	Colonoscopy: 10 years; FIT: Annually; FIT-DNA: Every 3-years	Supports emerging AI technologies in clinical research and potential guideline updates
American Gastroenterological Association	45	Colonoscopy preferred; other options acceptable for shared decision-making	Based on method	Draft guidelines recommend AI-assisted colonoscopy (CADe) as a quality enhancing tool
European Society of Gastrointestinal Endoscopy	50 (varies by country)	FIT-based programs, colonoscopy in select setting	FIT: Every two years; colonoscopy: Every ten years	Encourages integration of CADe systems in clinical practice to improve adenoma detection rates
National Comprehensive Cancer Network	45	Colonoscopy, stool-based tests, CT colonography	Colonoscopy: Ten years; FIT/FIT-DNA: Per test	Supports AI use for colonoscopy enhancement particularly in high-risk population
World Health Organization	50 (resource dependent)	FIT-based screening for population level programs	FIT: Every two years	Recognizes potential role of AI in expanding access and accuracy in low-resource setting; encourages cost-effective AI research
Japanese Society of Gastroenterology	40 (or even earlier for workplace health checks)	FIT	FIT: Annually	Japan’s first major AI-colonoscopy system, EndoBRAIN, received regulatory approval and reimbursement starting in 2024
Food and Drug Administration	45 (earlier initiation for those with higher risk factors)	Colonoscopy, FIT, stool DNA-FIT, CT colonography, flexible sigmoidoscopy	Colonoscopy: Every ten years; FIT: Annually; Stool DNA-FIT: Every 1-3 years; CT colonography: Every five years; Flexible sigmoidoscopy: Every 5-10 years	FDA has cleared CADe systems for real-time detection assistance during colonoscopy, but explicitly prohibits there use for lesion diagnosis or automated clinical decision making

AI: Artificial intelligence; FIT: Fecal immunochemical test; CT: Computed tomography; CADe: Computer-aided detection; FDA: Food and Drug Administration.