Performance of computer-aided diagnosis for colorectal sessile serrated lesions: A systematic review and meta-analysis

Abstract

BACKGROUND

The efficacy of computer-aided diagnosis (CADx) systems in identifying sessile serrated lesions (SSLs), which are critical precancerous lesions in colorectal cancer, remains unclear.

AIM

To comprehensively evaluate the diagnostic performance of CADx systems in differentiating between SSLs and non-SSLs and hyperplastic polyps (HPs).

METHODS

MEDLINE, EMBASE, and the Cochrane Library were searched up to 11 June 2025 for studies evaluating the performance of CADx systems in differentiating SSLs. The primary outcomes were the pooled diagnostic accuracy, sensitivity, and specificity of the CADx systems in distinguishing SSLs from non-SSLs or HPs.

RESULTS

Nine studies encompassing 2915 images and 746 videos on SSL differentiation were included. For SSLs vs non-SSLs, the CADx system demonstrated an overall area under the curve (AUC) of 0.93, 66% sensitivity, 95% specificity, a positive predictive value (PPV) of 0.56, a negative predictive value (NPV) of 0.96, a positive likelihood ratio (LR+) of 12.3, and a negative likelihood ratio (LR-) of 0.36. For SSLs vs HPs, the overall AUC was 0.64, with 55% sensitivity, 64% specificity, a PPV of 0.40, an NPV of 0.80, an LR+ of 1.5, and an LR- of 0.70. The sensitivity analysis indicated stable findings, whereas the latest World Health Organization pathological standards, image classification algorithms (ICAs), real-time scenarios, multicenter settings and narrow band imaging (NBI) significantly affected CADx system sensitivity. ICA served as an independent factor influencing the sensitivity of differentiating between SSLs and non-SSLs [odds ratio (OR) = 14.13], whereas NBI was an independent factor influencing the sensitivity of differentiating between SSLs and HPs (OR = 7.27) according to multivariate meta-regression.

CONCLUSION

Current CADx systems cannot adequately differentiate SSLs from non-SSLs or HPs. Future development should focus on improving the differentiation capability and sensitivity of SSLs.

Key Words: Computer-aided diagnosis; Sessile serrated lesions; Colonoscopy; Diagnostic accuracy; Sensitivity; Specificity

Core Tip: This is the first meta-analysis to comprehensively evaluate the efficacy of computer aided diagnosis (CADx) systems in identifying sessile serrated lesions (SSLs). The meta-analysis demonstrated the limited differentiation performance of CADx systems for SSLs, which is also significantly inferior to the distinguishing ability of colonoscopists. Although CADx systems perform well in differentiating SSLs from non-SSLs with relatively high specificity, their sensitivity is far lower than that of the optical biopsy criteria, particularly for differentiating between SSL and hyperplastic polyp.

INTRODUCTION

Timely identification of sessile serrated lesions (SSLs) is crucial for reducing the incidence of colorectal cancer (CRC)[1-4]. Approximately 8.6% of CRC cases are not detected promptly following colonoscopies[5], resulting in so-called postcolonoscopy CRC, where SSLs are considered one of the major contributing factors[1,4,6]. Notably, SSLs have malignant potential, particularly those with dysplasia, which can rapidly progress to invasive cancer[7,8]; however, not all colorectal polyps are at risk of developing into CRC. Among the screening or follow-up population, the polyp detection rate is as high as 60%, and only some of the lesions are adenomas or SSLs[9]. Diminutive polyps account for approximately 70%-80% of detected polyps, only 2% of which are found to have advanced histology[10,11]. Consequently, the resection of these low-risk polyps may unnecessarily increase the risk for complications, patient burdens, and medical costs.

The “resect and discard” and “leave-in situ” strategies aim to increase the cost-effectiveness of optical biopsies for accurate diagnosis and management of polyps, thereby reducing unnecessary polypectomies and pathological examination costs. The American Society for Gastrointestinal Endoscopy has established specific diagnostic performance standards for these strategies, requiring a 90% negative predictive value (NPV) for the “resect and discard” strategy and 90% concordance with surveillance intervals for the “leave-in situ” strategy[12]. Artificial intelligence (AI)-assisted computer-aided diagnosis (CADx) has been introduced as a potential solution for implementing optical biopsy strategies[13,14]. Several studies have demonstrated that CADx systems can reliably distinguish between adenomatous and nonadenomatous polyps[15-20], but their diagnostic performance for detecting SSLs remains unknown. Most studies have focused on the accuracy of optical diagnosis for adenomas, with SSLs excluded or labeled as nonneoplastic polyps[18,21], resulting in limited optical biopsy performance for SSLs.

Although several CADx systems have reported the ability to detect SSLs, the findings are variable and occasionally contradictory without a definitive conclusion[22-30]. Therefore, we conducted this systematic review and meta-analysis to comprehensively evaluate the performance of CADx systems in differentiating SSLs.

MATERIALS AND METHODS

This systematic review followed the preferred reporting items for systematic reviews and meta-analyses guidelines. The protocol is registered with PROSPERO (No. CRD420251060533).

Search strategy

A predefined comprehensive literature search was performed using core databases, including MEDLINE, EMBASE, and the Cochrane Library, up to 11 June 2025 to identify full articles evaluating the accuracy of CADx systems in SSL differentiation. The complete search strategy and search string can be found in the Supplementary material.

Inclusion and exclusion criteria

We included studies reporting computer-aided differentiation of colorectal SSLs, where histology served as the reference to define the accuracy of optical diagnosis, regardless of whether endoscopic images or videos were used. The performance of CADx systems in differentiating SSLs, including SSLs vs non-SSLs [including hyperplastic polyps (HPs), adenomas, adenocarcinomas, and other colorectal neoplasia] or SSLs vs HPs, should be able to estimate the true positive rate (TP), true negative rate (TN), false positive rate (FP), and false negative rate (FN). The exclusion criteria are described in the Supplementary material.

Selection process and quality assessment

Two reviewers independently screened the titles and abstracts of all identified articles to exclude those not related to the research topic or meeting one of the exclusion criteria and assessed the study quality using the Quality for Assessment of Diagnostic Studies 2 score[31]. For titles and abstracts deemed eligible, two reviewers independently reviewed the full reports to further determine eligibility for inclusion in our review. Any disagreements were resolved through discussions with the senior author (Zhao SB) until a consensus was reached.

Data extraction and analysis

Two reviewers independently extracted data from each eligible study and cross-validated the extracted data. Disagreements among the reviewers were resolved through discussion. If discussions were unsuccessful, disagreements in the extracted data were resolved by another independent arbitrator (Zhao SB). The TP, FP, FN, and TN values from each study were extracted, as well as other variables, such as publication, design, single-center or multicenter, geographical location, type of endoscope, histopathological standard, type of algorithm, real-time use, images- or video-based analysis, types of digital chromoendoscopy, and whether CADx was used as an assistive tool or alone. When one study reported data in different arms, we extracted the diagnostic accuracy data for each arm separately. The images or videos of the study should be independent and cannot have an inclusion or subordination relationship.

Statistical analysis

The pooled diagnostic accuracy, sensitivity, and specificity of the CADx systems in differentiating SSLs from non-SSLs or HPs were the main outcomes. The secondary outcomes included the pooled positive predictive value (PPV) and NPV, as well as the positive likelihood ratio (LR+), negative likelihood ratio (LR-), and diagnostic odds ratio (DOR). Both the primary and secondary outcomes were pooled using a bivariate random effects regression model. The accuracy of the AI was calculated as the area under the hierarchical summary receiver operating characteristic curve [area under the curve (AUC)], and Fagan’s plot is provided. LR+ and LR- were applied for the pooled probability of differentiating SSLs (i.e., the pretest probability) to calculate the posttest probability in the case of a positive or negative test result. If I² > 50%, a random effects model was employed to pool the outcomes. Subgroup analyses were performed on the basis of factors that may have contributed to heterogeneity. Bivariate boxplots were used to further investigate heterogeneity. Deeks’ funnel plot was drawn to examine the publication bias of the included studies. Sensitivity analysis was adopted to assess the robustness of the findings after individual studies were eliminated. Univariate and multivariate meta-regression analyses using mixed-effects models were conducted to explore predictive factors influencing the performance of CADx systems. The main analyses were performed using the midas package for Stata 15.0 and R (version 4.2.0). A two-sided P value < 0.05 was considered to indicate statistical significance.

RESULTS

Characteristics of the studies

The search strategy yielded a total of 5489 studies (Supplementary Tables 1-3). Once duplicates were removed, 3828 studies were screened by analysis of their titles and abstracts, and 3730 studies were removed because they were not related to the study topic or met one or more exclusion criteria. Afterward, 98 studies were further assessed for eligibility by full-text review, and 89 studies were excluded. Finally, 9 articles were included in the final analysis (Figure 1).

Figure 1  Flowchart of included studies.

Three studies were conducted in Europe, and six were conducted in Asia (Table 1). Four CADx systems were based on image classification algorithms (ICAs), whereas object detection algorithms (ODAs) were used in five studies[22,25-27,30]. Four studies used the latest fifth version of the World Health Organization (WHO) classification of tumors of the digestive system, whereas five studies used the fourth version. The total number of images in the test datasets was 2915, and that of videos was 746. Notably, 386 (319 images and 67 videos) out of 3661 were SSLs (including 302 lesions). The characteristics of the CADx system studies differentiating SSLs from non-SSLs are shown in Table 1, whereas those differentiating SSLs from HPs are illustrated in Supplementary Table 4.

Table 1 Study characteristics for sessile serrated lesions vs non-sessile serrated lesions, n (%).

Ref.	Design1	Center	Country	Type of endoscope	Polyp size	Histology criteria of WHO	AI segmentation/classification type	Real time	Images/videos	Imaging type	Image set (n)	Patients (n)	SSL image set (n)	SSL
Ozawa et al[22]	R	Single	Japan	Olympus	NA	2010	Single Shot MultiBox Detector/ODA	N	Images	WLI	783	174	23	7 (2)
Ozawa et al[22]	R	Single	Japan	Olympus	NA	2010	Single Shot MultiBox Detector/ODA	N	Images	NBI	290	174	6	7 (2)
Pu et al[24]	R	Single	Japan	Olympus	NA	2010	DenseNets/ICA	N	Images	NBI	20	NA	2	2 (10)
Pu et al[24]	R	Single	Japan	Fujifilm	NA	2010	DenseNets/ICA	N	Images	BLI	49	NA	10	10 (20)
Kader et al[28]	P	Single	United Kingdom	Olympus	NA	2019	Resnet101/ICA	N	Videos	NBI	1572	199	272	27 (17)
Kader et al[28]	P	Single	United Kingdom	Olympus	NA	2019	Resnet101/ICA	N	Videos	NBI	652	31	102	10 (15)
Houwen et al[27]	P	Multiple	Netherlands	Olympus	≤ 5	2010	YOLO v4/ODA	Y	Videos	NBI	4232	194	72	7 (2)
Song et al[23]	R	Single	Korea	Olympus	NA	2010	DenseNet-201/ICA	N	Images	NBI	182	182	39	24 (13)
Song et al[23]	P	Single	Korea	Olympus	NA	2010	DenseNet-201/ICA	N	Images	NBI	546	363	117	70 (19)

¹Represents a retrospective or prospective study.

²Represents the use of video data.

AI: Artificial intelligence; WHO: World Health Organization; SSL: Sessile serrated lesion; R: Retrospective; P: Prospective; NA: Not available; ODA: Object detection algorithm; ICA: Image classification algorithm; N: No; Y: Yes; WLI: White light images; NBI: Narrow band imaging; BLI: Blue light imaging.

Study quality and publication bias

The results of the study quality assessment are presented in Supplementary Table 5, and all included studies were considered high quality (Supplementary Figure 1). According to Deeks’ regression test (P = 0.96 and 0.65, respectively), no significant publication bias was detected for the comparison of SSLs vs non-SSLs or for the comparison of SSLs vs HPs (Supplementary Figure 2).

Performance of CADx systems in differentiating between SSLs and non-SSLs

Data on the performance of CADx systems in differentiating SSLs from non-SSLs were available in 5 studies (9 arms)[22-24,27,28], including a total of 157 SSLs (214 related images and 44 videos).

In terms of accuracy, the AUC for the use of CADx in differentiating SSLs from non-SSLs was 0.93 [95% confidence interval (CI): 0.91-0.95; Figure 2A], whereas the other diagnostic metrics yielded a pooled sensitivity of 66% (95%CI: 35%-87%; Figure 2B), a specificity of 95% (95%CI: 87%-98%; Figure 2B), a PPV of 56% (95%CI: 0.37-0.75; Supplementary Figure 3A), an NPV of 96% (95%CI: 94%-98%; Supplementary Figure 3B), an LR+ of 12.3 (95%CI: 6.2-24.4; Supplementary Figure 4), an LR- of 0.36 (95%CI: 0.16-0.81; Supplementary Figure 4), and a DOR of 34 (95%CI: 13-85; Supplementary Figure 5). Heterogeneity was observed for both sensitivity (I² = 91.08; Figure 2B) and specificity (I² = 93.95; Figure 2B). In absolute terms, for a 20% prevalence of SSLs, a positive result from the CADx system increased the probability of disease to 75%, whereas a negative result decreased the probability of disease to 8% (Supplementary Figure 6A). Additional analyses, assuming an SSL prevalence of 50%, were also conducted (Supplementary Figure 6B).

Figure 2 Receiver operating characteristic curve and forest plots of sensitivity and specificity. A and B: In differentiating sessile serrated lesions (SSLs) from non-SSLs; A: Summary receiver operating characteristic curve for differentiating SSLs from non-SSLs; B: Forest plots of sensitivity and specificity in differentiating SSLs from non-SSLs; C and D: In differentiating SSLs from hyperplastic polyps (HPs); C: Summary receiver operating characteristic curve for differentiating SSLs from HPs; D: Forest plots of sensitivity and specificity in differentiating SSLs from HPs. SROC: Summary receiver operating characteristic; SENS: Sensitivity; SPEC: Specificity; AUC: Area under the curve; CI: Confidence interval.

Performance of CADx systems in differentiating SSLs from HPs

Data on the performance of CADx systems in differentiating SSLs from HPs were available in 7 studies (10 arms)[22,24-26,28-30], including a total of 201 SSLs (163 images and 60 videos).

In terms of accuracy, the AUC for the use of CADx in differentiating SSLs from HPs was 0.64 (95%CI: 0.60-0.68; Figure 2C), whereas other diagnostic metrics yielded a pooled sensitivity of 55% (95%CI: 21%-85%; Figure 2D), a specificity of 64% (95%CI: 41%-82%; Figure 2D), a PPV of 40% (95%CI: 0.19-0.61; Supplementary Figure 7A), an NPV of 80% (95%CI: 0.71-0.89; Supplementary Figure 7B), an LR+ of 1.5 (95%CI: 0.6-3.8; Supplementary Figure 8), an LR- of 0.70 (95%CI: 0.28-1.73; Supplementary Figure 8), and a DOR of 2 (95%CI: 0-13; Supplementary Figure 9). Heterogeneity was observed for both sensitivity (I² = 91.22; Figure 2D) and specificity (I² = 91.58; Figure 2D). In absolute terms, for a 20% prevalence of SSLs, a positive result from CADx increased the probability of disease to 28%, whereas a negative result decreased the probability of disease to 15% (Supplementary Figure 10A). Additional analyses according to the hypothesized pretest prevalence of SSLs of 50% are also provided (Supplementary Figure 10B).

Ability of endoscopists to differentiate SSLs from non-SSLs/HPs

Two studies (3 arms) reported the ability of endoscopists to differentiate SSLs from non-SSLs, and two other studies reported the ability of endoscopists to differentiate SSLs from HPs. The pooled analysis revealed that in distinguishing SSLs from non-SSLs, the combined sensitivity, specificity and total accuracy of endoscopists were 79% (95%CI: 0.64-0.89; Supplementary Figure 11A), 92% (95%CI: 0.90-0.94; Supplementary Figure 11B), and 86% (95%CI: 0.83-0.88; Supplementary Figure 11C), respectively. However, for differentiating SSLs from HPs, the combined sensitivity, specificity and total accuracy of endoscopists were 72% (95%CI: 0.55-0.85; Supplementary Figure 12A), 71% (95%CI: 0.60-0.79; Supplementary Figure 12B) and 71% (95%CI: 0.66-0.75; Supplementary Figure 12C), respectively. Within the same study, the CADx system demonstrated a non-significantly lower sensitivity compared with the endoscopists in all comparisons.

Subgroup and sensitivity analyses

Compared with the SSL vs non-SSL groups, single-center studies (P = 0.02) and use of the WHO pathological standard (P = 0.01), ICAs (P < 0.001), and nonreal-time settings significantly increased the sensitivity (P = 0.02; Figure 3A). The use of ICAs also yielded higher specificity (P < 0.001; Figure 3A).

Figure 3 Forest plots of subgroup analysis. A: Sessile serrated lesions (SSLs) vs non-SSLs; B: SSLs vs hyperplastic polyps. ^aP < 0.05. ^bP < 0.01. ^cP < 0.001. CI: Confidence interval; pt: Publication time; ICA: Image classification algorithms; NBI: Narrow band imaging.

Compared with the SSLs vs HPs group, single-center studies (P < 0.001) and the use of ICAs (P = 0.02) and narrow band imaging (NBI) (P = 0.01) significantly increased the sensitivity (Figure 3B). The use of videos yielded higher specificity (P = 0.04; Figure 3B). No significant effect on the sensitivity or specificity was found for the other prespecified variables (Supplementary Table 6).

To further investigate heterogeneity, a bivariate boxplot was used to demonstrate the degree of interdependence of sensitivity and specificity, including the identification of potential outliers (Supplementary Figure 13). After each study (including the outliers) was eliminated, we recalculated the sensitivity, specificity, LR+, LR- and DOR values, and the sensitivity analysis indicated that the results remained stable (Supplementary Table 7).

Meta-regression for the sensitivity and specificity of distinguishing SSLs from non-SSLs or HPs

In the differentiation of SSLs vs non-SSLs using CADx systems, univariate regression of sensitivity revealed that ICAs (OR = 16.22, 95%CI: 7.73-34.05; P < 0.01) and the 2019 criteria (OR = 11.15, 95%CI: 0.90-138.81; P = 0.06) significantly improved sensitivity, whereas only the use of an ICA (OR = 14.13, 95%CI: 6.89-29.01; P < 0.01) remained an independent factor in the multivariate regression (Table 2). According to the results of the specificity analysis, only the use of an ICA significantly reduced specificity (univariate OR = 0.16, 95%CI: 0.06-0.42; P < 0.01) (Supplementary Table 8).

Table 2 Meta regression analysis for possible factors affecting the sensitivity of computer-aided diagnosis in distinguishing sessile serrated lesions from non-sessile serrated lesions and hyperplastic polyps.

Factors	Sensitivity (SSL vs non-SSLs)				Sensitivity (SSL vs HPs)
	Univariate analysis		Multivariate analysis		Univariate analysis		Multivariate analysis
	P value	OR (95%CI)	P value	OR (95%CI)	P value	OR (95%CI)	P value	OR (95%CI)
Publication time (vs year 2020)	0.67	1.77 (0.13-23.36)			0.48	2.49 (0.20-31.49)
Multicenter (vs single center)	0.10	0.09 (0.01-1.61)			0.06	0.03 (0.00-1.19)	0.14	0.10 (0.00-2.09)
Europe (vs Asia)	0.67	1.77 (0.13-23.36)			0.64	1.97 (0.12-33.01)
Prospective (vs retrospective)	0.67	1.77 (0.13-23.36)			0.33	3.40 (0.29-39.44)
Olympus (vs Fujifilm)	0.83	0.66 (0.02-27.83)			0.11	6.84 (0.65-72.26)
Criteria 2019 (vs 2010)	0.06	11.15 (0.90-138.8)	0.07	2.97 (0.92-9.53)	0.19	4.55 (0.47-43.74)
ICA (vs ODA)	< 0.01	16.22 (7.73-34.05)	< 0.01	14.13 (6.89-29.01)	0.18	5.28 (0.47-59.22)
Real time (vs nonreal time)	0.10	0.09 (0.01-1.61)			0.58	0.41 (0.02-9.53)
Images (vs videos)	0.67	0.57 (0.04-7.47)			0.64	0.51 (0.03-8.48)
NBI (vs non-NBI)	0.47	2.68 (0.18-39.61)			0.01	10.80 (1.97-59.33)	0.02	7.27 (1.40-37.63)

SSL: Sessile serrated lesion; HPs: Hyperplastic polyps; OR: Odds ratio; CI: Confidence interval; ICA: Image classification algorithms; ODA: Object detection algorithm; NBI: Narrow band imaging.

With respect to the differentiation of SSLs vs HPs by CADx systems, according to the sensitivity analysis, only NBI significantly improved the sensitivity in both univariate (OR = 10.80, 95%CI: 1.97-59.33; P = 0.01) and multivariate analyses (OR = 7.27, 95%CI: 1.40-37.63; P = 0.02). With respect to specificity, recent studies (2023-2024) significantly increased specificity (univariate OR = 3.91, 95%CI: 1.27-12.09, P = 0.02). In contrast, European studies (OR = 3.82, 95%CI: 0.98-14.99, P > 0.05), the 2019 criteria (OR = 3.63, 95%CI: 0.98-13.52, P > 0.05), and image type (OR = 0.26, 95%CI: 0.07-1.03, P > 0.05) were marginally significant, with no factors identified in the multivariate regression (Supplementary Table 8).

DISCUSSION

Our meta-analysis demonstrated the limited ability of CADx systems to detect SSLs. Notably, although CADx systems perform well in differentiating SSLs from non-SSLs with relatively high specificity, their sensitivity is far lower than that of the optical biopsy criteria, particularly for differentiating between SSL and HP. Further efforts should be focused on the use of high-quality SSL videos/images and state-of-the-art classification algorithms to optimally detect SSLs.

Serrated lesions exhibit distinct genetic characteristics, including a high presence of the BRAF V600E mutation and activation of the mitogen-activated protein kinase signaling pathway, which are strongly associated with the CpG island methylator phenotype and the presence of microsatellite instability, all of which are important for understanding the diagnosis, treatment, and prevention of CRC progression through the serrated neoplasia pathway. Although endoscopic diagnosis for SSL was suboptimal, this study reveals the key bottlenecks of the CADx system in distinguishing SSLs. The CADx system has relatively low sensitivity and PPV but high specificity and NPV in differentiating SSLs from non-SSLs (mainly adenomas). This is primarily explained by two factors. First, although SSLs and non-SSLs (particularly adenomas) exhibit some distinguishable morphological features, allowing the CADx system to achieve high specificity (correctly excluding non-SSLs from adenomas) by recognizing specific structural patterns, SSLs often present as flat and cloud-like lesions with overlying mucus[32]. This makes their static imaging features easily confusable with those of HPs, leading to insufficient sensitivity (missed SSL diagnoses). Second, the overall prevalence of SSLs among colorectal lesions is low, which further reduces the PPV, while the high specificity supports the NPV.

However, the differentiation of SSLs from HPs faces greater challenges: Both belong to the serrated lesion spectrum with greater morphological similarity typically presenting as flat/slightly elevated lesions with indistinct surface patterns. This makes CADx systems prone to miss subtle features, yielding reductions in both sensitivity and specificity. Furthermore, HP overrepresentation in training data biases the model toward HP misclassification (increasing SSL FPs), further decreasing the PPV. The acceptable NPV reflects the inherently low malignant potential of HPs, making FNs clinically less consequential. This performance falls short of the dual preservation and incorporation of valuable endoscopic innovations-90% benchmarks (requiring both sensitivity and NPV ≥ 90%), carrying two clinical risks. First, low sensitivity may lead to underdiagnosis of SSLs (misdiagnosed as HPs), which may increase interval cancer risk through the serrated neoplasia pathway. Second, low specificity means shorter surveillance intervals through misidentification of nonneoplasms (such as HPs) as neoplasms (such as SSLs), resulting in unnecessary endoscopic removal of HPs, increased patient medical costs and inefficient use of medical resources. Both the high FP rate and FN rate demonstrated the poor performance of the CADx system, which may lead to an elevated risk in clinical decision-making and should not be used in clinical practice. Furthermore, the low performance of CADx systems in differentiating SSLs from other lesions may be attributed to a lack of pretraining. One possible explanation for the poor performance of CADx systems could be the lack of pretraining of the models, which could be improved by increasing the quantity of data but, more importantly, by improving the variety and quality of the data.

NBI is currently among the most widely used electronic staining techniques. Studies have confirmed that when non-adenomatous polyps are diagnosed endoscopically, NBI is superior to white light endoscopy[33]. Under NBI, the appearance of the SSL crypts is cloud-like, with an irregular shape, black dots inside the crypts, and type II pit pattern glandular duct openings. NBI may help detect more proximal colonic serrated lesions (0.51 vs 0.39, P = 0.085)[34]. Our meta-analysis also confirmed that the typical characteristics of SSLs are more easily identifiable using NBI compared with other techniques. NBI demonstrated specialized efficacy in improving sensitivity for SSL discrimination vs HP discrimination (OR = 7.27). Applying the latest diagnostic criteria may increase the sensitivity for diagnosing SSLs and reduce the error rate. In real-time settings, owing to the complex environment, which includes factors such as operation, imaging, and onsite conditions, the sensitivity is relatively low.

Both the subgroup analysis and meta-regression analyses revealed that ICAs significantly enhance the sensitivity (P < 0.001, OR = 14.13) and specificity (P < 0.001, OR = 0.16) of SSL detection vs non-SSL detection compared with ODAs, with low heterogeneity noted in each subgroup. ODAs have been widely applied in the computer-aided detection of colorectal lesions and have difficulty differentiating between SSL and other lesions[29], whereas ICAs classify colorectal lesions into at most five categories with high AUCs[24]. The results provided by subgroup analyses can serve as a reference for future research; head-to-head studies, especially randomized controlled trials, should be conducted to further confirm these findings.

SSLs are often indistinguishable from HPs but are classified as neoplastic given their malignant potential. However, various previous CADx studies excluded SSLs or treated them as nonneoplasms[16,17,21]. Considering the resect-and-discard strategy, SSLs should be precisely diagnosed as it was one of the CRC precursors and are included in surveillance guidelines[35,36]. However, the performance of CADx systems in distinguishing SSLs is unsatisfactory. The difficulty in developing a CADx system capable of differentiating SSLs might be caused by interobserver variability among pathologists[37]. Since the concept of serrated lesions was proposed, their diagnostic terms and pathological diagnostic criteria have undergone multiple updates. In the past, they were called serrated adenomas and sessile serrated adenomas/polyps, and these diagnostic terms have sparked numerous disputes among pathologists[38,39]. By 2019, the 5^th WHO classification proposed the diagnostic term “sessile serrated lesion” and stated that SSLs can be diagnosed if characteristic structural changes occur in at least one crypt[36]. Because these morphological features partly overlap with those of HPs, distinguishing between the two can be challenging. The changes in the histological criteria for SSLs, where some micro-vesicular-type HPs (MVHPs) are reclassified into SSLs, indicate that SSLs may not be a disease entity distinct from HPs. This leads to significant controversies among pathologists in the diagnosis of SSLs, with poor consistency and high variability in histopathological diagnoses[40]. The kappa value from the overall consistency assessment for SSL diagnosis by pathologists in Europe and the United States is only 0.44, with moderate performance[41]. In clinical practice, the presence of a lesion > 10 mm in the proximal colon is an indication for polypectomy, regardless of histology. European Society of Gastrointestinal Endoscopy recommends that all SSLs of any size without dysplasia be resected by a cold snare or cold endoscopic mucosal resection[42]. In addition to SSLs, small (< 10 mm) serrated lesions are not ideal candidates for endoscopic resection, highlighting the clinical importance of detecting and diagnosing SSLs. The orientation of the lesions used in the development of CADx systems was regarded as an important influencing factor for system performance irrespective of adequate visualization to the human eye[28]. Another plausible explanation is that most systems are developed using data from a small number and proportion of SSL pictures with limited orientation, which limits the full use of deep neural networks to extract features for classification and may influence their efficacy in distinguishing SSLs. Considering that direct endoscopic differentiation between MVHP and SSL remains challenging at present, future studies should investigate the ability of endoscopist features to distinguish between SSLs and MVHPs. Furthermore, SSLs with dysplasia (SSLD), characterized by rapid transformation of SSL into CRC, is crucial for the progression to invasive carcinoma. Although the studies included did not report the sensitivity and specificity of CADx in patients with SSLD, evaluating the overall performance of CADx was not possible. Therefore, future studies with larger sample sizes are warranted to focus on the performance of CADx in the diagnosis of SSLD, and further meta-analysis is needed to evaluate the overall effect.

Furthermore, several limitations prevent the implementation of CADx systems in clinical practice. First, most models were developed on the basis of deep learning with opaque decision-making processes, which can raise concerns about their reliability and trustworthiness among clinicians and patients. Recent studies have enabled endoscopists to directly identify regions with a high probability of lesions through the red-highlighted areas displayed on the monitor[24,25], which largely improved the interpretability of the model. However, further attempts are needed. Second, the vast majority of the studies were developed and validated with a limited number of high-quality still images after washing and suctioning rather than videos. When bowel preparation is inadequate, factors such as intestinal mucus, fecal residue, peristalsis, motion artifacts, and variable illumination can substantially affect lesion visibility and image quality, which was not considered during model development and may result in poor performance of CADx systems[43]. In addition, the images were extracted with predefined inclusion/exclusion criteria and subjective assessments of image quality, which also generalize less well to nonexpert endoscopists, who may be less likely to reproduce the high-quality images of experts[28]. In summary, models developed from only images have difficulty capturing dynamic factors such as intestinal mucus, thus limiting their generalizability to clinical practice. Further CADx models should be developed and validated using videos rather than images to improve the performance and dissemination of CADx in clinical practice. Third, endoscopist experience was an important factor in CADx application, as experts achieved improvements in all evaluation indicators after the use of CADx[30], indicating that experienced endoscopists rather than trainees may further improve diagnostic precision when supported by CADx systems. Fourth, a recent analysis revealed that the diagnostic accuracy of endoscopists may decrease following long-term use of AI-assisted endoscopy[44]. These findings also emphasize the necessity of guarding against the potential blunting effect exerted by such AI on endoscopists.

Some limitations still exist. First, most of the included studies were based on pictures, which limits comparisons of the accuracy of CAD systems between pictures and videos and subgroup analyses based on bowel preparation quality. Second, few studies included in the current meta-analysis reported the efficacy of CADx in distinguishing between SSLD and MVHP; thus, an evaluation of the performance of CADx on SSLD/MVHP and an investigation of potential independent factors were not possible. Third, the included studies involved a single colonoscopy rather than a tandem design with no follow-up data, and it is not possible to obtain a specific miss rate for SSLs. Thus, we cannot evaluate the performance of CADx for missed SSLs. Finally, the number of eligible studies was not large, and the vast majority of the included studies (7/9) did not use the SSL as the primary endpoint. This may have led to statistical bias and inaccurate evaluation of the performance of the CADx models.

CONCLUSION

In conclusion, our analysis revealed poor CADx performance for the differentiation of SSLs, especially for differentiating between SSLs and HPs. This is a key field that remains in its infancy and is likely to make progress using deep learning techniques, as CADx systems may be able to improve optical diagnosis performance and reduce colonoscopy costs.