Low accuracy and interobserver/intraobserver agreement for colorectal sessile serrated lesions: A multicenter diagnostic study

Abstract

BACKGROUND

The accurate diagnosis of sessile serrated lesions (SSLs) and SSLs with dysplasia (SSL-D) is vital for preventing the occurrence of colorectal cancer but remains challenging because of inconsistent classification criteria.

AIM

To evaluate the diagnostic consistency for colorectal SSL and SSL-D among pathologists with different levels of experience in a multicenter setting.

METHODS

Six gastrointestinal pathologists (stratified by experience) independently reviewed digitized whole-slide images of 1176 colorectal polyps across three rounds using the 2019 World Health Organization criteria. The accuracy and interobserver/intraobserver agreement were assessed. Logistic regression analysis was performed to identify factors influencing SSL/SSL-D concordance.

RESULTS

Overall diagnostic accuracy increased from 66% to 72% with clinical information (P < 0.001), but accuracy for SSL (62%) and SSL-D (33%) remained suboptimal. Interobserver agreement was moderate for SSL (κ = 0.301-0.441), whereas the highest agreement was observed for traditional serrated adenoma (κ = 0.703). Senior pathologists outperformed junior pathologists in terms of the overall (accuracy: 76% vs 56%, P < 0.001; κ: 0.517 vs 0.337) and SSL classification (69% vs 60%, P < 0.001; κ: 0.601 vs 0.524). SSL-D showed the poorest performance regardless of experience (accuracy: 11%-21%; interobserver κ: 0.084-0.149). Multivariate analysis identified nuclear alteration [odds ratio (OR) = 10.45)] and cross-sectional crypt orientation (OR = 5.27) as risk factors for discordance, whereas > 3 serrated crypts (OR = 0.37) and serration extending to the crypt base (OR = 0.20) exhibited protective effects.

CONCLUSION

Suboptimal performance and high diagnostic variability highlight the morphological complexity of SSL/SSL-D classification. Standardized training and quantitative crypt-based criteria may improve diagnosis.

Key Words: Colorectal polyps; Sessile serrated lesion; Pathological diagnosis; Diagnostic consistency; Histopathological assessment

Core Tip: In this multicenter diagnostic study of 1176 colorectal lesions, pathologists’ accuracy and agreement for sessile serrated lesions (SSLs) and SSLs with dysplasia were evaluated. After clinical information was provided, the overall diagnostic accuracy improved from 66% to 72%; however, the diagnostic consistency of SSLs with dysplasia remained poor (κ = 0.114-0.132). Nuclear alterations (odds ratio = 10.45) and the cross-sectional crypt orientation (odds ratio = 5.27) were identified as factors that influenced diagnostic inconsistency. These results suggest that experience is insufficient to overcome the diagnostic challenges of SSLs. This finding reinforces the need for quantitative standards, algorithms for quality control of section orientation, and artificial intelligence-assisted curative devices to improve diagnostic precision.

INTRODUCTION

Colorectal cancer (CRC) is one of the most common cancers worldwide. Its incidence and mortality have consistently been ranked among the top three cancers for both sexes[1]. The cornerstone of all CRC screening programs is the early endoscopic resection of premalignant lesions and an accurate histopathological diagnosis. Colorectal adenomas are clear precancerous lesions that have been recognized and demonstrated as effective targets for any endoscopic resection with risk-stratified follow-up to prevent CRC[2,3]. However, many people develop interval cancer within five or ten years after endoscopy. Sessile serrated lesions (SSLs) are thought to be major cause of interval cancer. Epidemiological studies have shown that 15%-30% of cases of CRC are due to serrated lesions[4-6]. Nevertheless, the endoscopic detection rate of SSLs is still low, between 4% and 6%, in screening cohorts[7-9].

Serrated lesions include various types of colorectal epithelial proliferation characterized histologically by sawtooth-like architectural features, ranging from benign hyperplastic polyps (HPs) to premalignant SSLs and SSLs with dysplasia (SSL-D). These lesions exhibit morphological heterogeneity. The variations can range from mild serrations confined to the superficial epithelium to complex branching patterns. Furthermore, the lesions also display cytological atypia[10]. Interobserver variability in the histopathological diagnosis of SSLs and HPs has been observed in routine practice. Furthermore, a diagnosis is difficult because of the overlapping morphology in many instances. The diagnosis of SSLs and SSL-D is particularly difficult because of their distinctive but subtle features, such as basal crypt dilatation and asymmetric maturation[11,12]. These features affect the detection of serrated lesions and the efficacy of colonoscopy in preventing CRC. However, a thorough assessment of the accuracy and reliability of the SSL diagnosis is lacking, and the factors affecting the histological classification of SSLs are unclear.

To fill these gaps, we conducted a multicenter, large-scale diagnostic investigation to determine the accuracy, interobserver agreement, and intraobserver agreement for determining the presence of colorectal polyps by pathologists, specifically SSLs and SSL-D.

MATERIALS AND METHODS

Case selection and data source

In this multicenter retrospective study, 1176 colorectal polyp samples collected from four tertiary medical institutions were analyzed. All the samples were histologically classified according to the World Health Organization (WHO) Classification of Digestive System Tumors (5^th edition, 2019), which includes two main categories: Serrated lesions [goblet cell-rich HPs, microvesicular HPs (MVHPs), SSLs, SSL-D, and traditional serrated adenomas (TSAs)] and conventional adenomas [tubular adenomas (TAs), villous adenomas, and tubulovillous adenomas, Figure 1)]. Samples were included only if all of the following histological criteria were met: Hematoxylin and eosin (HE)-stained 4-5-μm-thick sections should display > 80% of the crypt architecture and be free of significant histological processing artifacts (e.g., crushing/fragmentation). The exclusion criteria were as follows: (1) Patients with CRC, inflammatory bowel disease, hamartomatous polyposis, and familial adenomatous polyposis; (2) Any samples with substandard technical quality (including insufficient materials or damaged slides); and (3) Lesion samples that remained undiagnosed after a consensus discussion. Ethical approval was obtained from the Ethics Committee of Changhai Hospital (No. CHEC2024-421), and the requirement for informed consent was waived because of the retrospective nature of this study and the use of anonymized pathological specimens.

Figure 1 Pathomorphological features of different types of colorectal polyp lesions. A: Similar to normal crypt morphology, but primarily composed of goblet cells with small, uniform nuclei located at the base (goblet cell-rich hyperplastic polyp, hematoxylin-eosin, original magnification 100 ×); B: The surface of the crypts exhibits serrated architectures, with the cross-section displaying a star-like pattern, and small droplet-like mucus can be observed scattered within the epithelial cells (microvesicular hyperplastic polyp, hematoxylin-eosin, original magnification 40 ×); C: The serrated crypts are prominently displayed, extending deeply into the crypts with dilation and branching, giving the gland a distinctive ‘L’-shaped configuration (sessile serrated lesion, hematoxylin-eosin, original magnification 100 ×); D: Serrated or fissure-like crypts are observed, with the presence of ectopic crypts. The nuclei exhibit elongation resembling a pencil, and the cytoplasm displays significant eosinophilia (traditional serrated adenoma, hematoxylin-eosin, original magnification 100 ×); E: The glandular tissue exhibits an irregular tubular pattern with a crowded arrangement of cells and deeply stained nuclei (tubular adenoma, hematoxylin-eosin, original magnification 100 ×); F: The glandular tissue exhibits an irregular tubular and villous configuration, with cells demonstrating atypia and deeply stained nuclei (tubulovillous adenoma, hematoxylin-eosin, original magnification 100 ×).

Whole-slide imaging

A Hamamatsu NanoZoomer S60 scanner (40 × objective magnification) was used to digitize all HE stained slides. The digital slides were stored in files in a proprietary format and uploaded to a password-protected server for remote access (Figure 2).

Figure 2  Research process workflow.

Diagnostic evaluation protocol

Pathologist composition: The diagnostic team included six gastrointestinal pathologists (Doc1-Doc6), who were divided into three groups based on the experience: (1) Senior pathologists (Doc1 and Doc2, ≥ 5 years of subspecialty experience); (2) Intermediate pathologists (Doc3 and Doc4, 3-5 years of experience); and (3) Junior pathologists (Doc5 and Doc6, < 3 years of experience, Figure 2).

Diagnostic protocol: Three sequential rounds of whole-slide image evaluation were conducted through a web-based platform (Figure 2). In every round, the pathologists independently analyzed the lesion morphology (serration pattern, crypt architecture, and cytological atypia) using the 2019 WHO criteria. Each evaluation phase was separated by a 4-week washout period to minimize recall bias. In the final round, anonymized clinical data [patient’s age, anatomical location of the lesion, lesion size, and endoscopic morphology (sessile or nonsessile)] were provided to the pathologists to mimic real-world conditions for diagnostic assistance. After three rounds of independent assessments, panel discussions were held for cases with diagnostic inconsistencies, with the majority opinion serving as the diagnostic gold standard.

Statistical analysis

Diagnostic accuracy, including both overall accuracy and lesion-specific accuracy, was determined with a consensus diagnosis as the gold standard. McNemar’s χ² test was used to evaluate the accuracy between rounds 1 and 3 and across pathologists of different experience levels for each round. Differences in lesion-specific accuracy were evaluated using Pearson’s χ² test in the first round. Interobserver agreement was assessed by calculating Fleiss’ κ with 95%CI, and the results were stratified by pathologist experience and lesion subtype. Intraobserver agreement was evaluated using Cohen’s κ for pairwise comparisons between the first and second rounds, with the Landis and Koch criteria used to define the agreement strength as follows: (1) κ < 0.20 slight agreement; (2) 0.21-0.40 fair agreement; (3) 0.41-0.60 moderate agreement; (4) 0.61-0.80 substantial agreement; and (5) > 0.81 almost perfect agreement[13]. For multiple comparisons of lesion-specific accuracies vs overall accuracy in round 1, the Bonferroni correction was applied to control the familywise error rate; the corrected significance threshold was set at P < 0.0071 (α = 0.05 divided by 7 lesion subtypes). Comparisons between round 3 and round 1 for each subtype were prespecified single contrasts and were not corrected, and a two-sided P < 0.05 was considered to indicate statistical significance. All the statistical tests were two-sided.

For the logistic regression analysis, SSL/SSL-D cases were stratified into a consensus group (agreement among ≥ 5/6 pathologists) and a discordance group to identify risk factors for diagnostic discordance. Demographic (age and sex), endoscopic (morphology and size), and histopathological (crypt orientation, serrated region, crypt dilation, proliferative zone, abnormal crypt architecture, number of serrated crypts, nuclear alteration and eosinophilia) variables were assessed. Variables with P < 0.1 in the univariate analysis were entered into the multivariate model. Collinearity was assessed by calculating the variance inflation factor (VIF), with a VIF > 5 exclusion threshold applied. Significant collinearity was observed among the serrated region, crypt dilation, and abnormal crypt architecture; the serrated region was retained based on its pathological significance. The results are reported as odds ratios (ORs) with 95%CIs. Statistical analyses were performed using SPSS 26.0 (IBM Corp.) and R 4.4.2.

RESULTS

Clinical characteristics of the patients

The study cohort included 1176 patients with histologically confirmed colorectal polyps. The mean age of the patients was 57.32 ± 12.00 years (range: 38-92 years), and most of the patients were female (68.4%). Endoscopic evaluations revealed that 27.0% of the lesions had a sessile morphology, whereas most of the lesions had a nonsessile morphology (73.0%). The analysis of the anatomical distribution revealed that the transverse colon (22.4%) and sigmoid colon (22.7%) were the most common locations, followed by the ascending colon (16.2%) and rectum (13.4%). Stratification by lesion size revealed that 35.2% of the polyps were less than 0.5 cm in size, 36.3% were between 0.5 and 1 cm, and 11.3% were more than 1 cm in size (Table 1).

Table 1 Clinical characteristics of the included cases, mean ± SD/n (%).

Parameters		Number of cases
Age (years)		57.32 ± 12.00
Sex	Female	804 (68.4)
	Male	372 (31.6)
Flat or sessile morphology	No	859 (73.0)
	Yes	317 (27.0)
Anatomical location	Cecum	72 (6.1)
	Ascending colon	191 (16.2)
	Hepatic flexure	82 (7.0)
	Transverse colon	264 (22.4)
	Splenic flexure	7 (0.6)
	Descending colon	136 (11.6)
	Sigmoid colon	267 (22.7)
	Rectum	157 (13.4)
Lesion size1	≤ 0.5 cm	414 (35.2)
	0.5 < x ≤ 1 cm	427 (36.3)
	> 1 cm	133 (11.3)

¹Polyp size data of some cases are missing.

Diagnostic accuracy among the pathologists

Diagnostic accuracy progressively increased across the three evaluation rounds, with the overall concordance among the pathologists increasing from 66% in round 1 to 72% in round 3 (P < 0.001; Table 2). Subtype-specific analyses revealed distinct patterns: The diagnostic accuracy for the SSLs decreased from round 1 to round 3 (62%-58%, P = 0.032), whereas that for the SSL-D increased (33%-49%, P = 0.007). Additionally, both rates were lower than the overall accuracy rate. In contrast, the accuracy for TAs and TSAs was greater than the cohort average, with the greatest absolute improvement from round 1 to round 3 achieved for TAs (66%-80%, P < 0.001) and high accuracy maintained for TSAs (74% vs 77%, P = 0.729; Table 2). Compared with the overall baseline accuracy (66%) in round 1, significantly lower accuracy was observed for SSL-D (33%, P < 0.001), whereas significantly higher accuracy was observed for goblet cell-rich HPs (71%, P < 0.001) and TSAs (74%, P = 0.004) after the Bonferroni correction for multiple comparisons (corrected threshold P < 0.0071). The differences in accuracies between round 1 and round 3 for the SSLs (62%, P = 0.014) and MVHPs (62%, P = 0.023) did not reach statistical significance after the Bonferroni correction, whereas the accuracies for the TAs (66%, P = 0.987) and villous adenomas/tubulovillous adenomas (65%, P = 0.815) remained comparable (Table 2). The results of the subgroup analysis stratified by expertise level demonstrated that, compared with junior pathologists, senior pathologists maintained significantly superior diagnostic accuracy across all the domains (P < 0.001), except for SSL-D assessments (Figure 3A). Intermediate pathologists outperformed junior pathologists in the later SSL-D rounds (round 2: 50% vs 18%, P = 0.012; round 3: 54% vs 11%, P < 0.001) and in overall accuracy (P < 0.001) but showed no significant advantage in the SSL assessments (Figure 3A).

Figure 3 Diagnostic accuracy and interobserver agreement by experience level. A: Diagnostic accuracy of different experience level and pathology subtype across reading rounds. Significance vs junior group: ^bP < 0.01, ^cP < 0.001; B: Interobserver agreement by experience level across reading rounds. Reference lines: κ = 0 (dashed), κ = 0.2/0.4 (dotted). ^aNegative values indicate agreement worse than chance. SSL: Sessile serrated lesion; SSL-D: Sessile serrated lesion with dysplasia.

Table 2 Diagnostic accuracy and interobserver agreement across reading rounds.

Subtype	Round 1			Round 2			Round 3			P accuracy (round 1 vs round 3)	P accuracy (vs overall in round 1)
	Accuracy	κ	95%CI	Accuracy	κ	95%CI	Accuracy	κ	95%CI
Overall	66%	0.393	0.385-0.400	69%	0.441	0.434-0.449	72%	0.415	0.407-0.423	< 0.001a	reference
SSL	62%	0.409	0.394-0.423	60%	0.441	0.426-0.455	58%	0.301	0.286-0.316	0.032a	0.014
SSL-D	33%	0.114	0.099-0.129	33%	0.132	0.117-0.147	49%	0.127	0.112-0.141	0.007a	< 0.001
GCHP	71%	0.369	0.354-0.384	72%	0.402	0.387-0.417	69%	0.369	0.355-0.384	0.102	< 0.001
MVHP	62%	0.332	0.318-0.347	68%	0.425	0.411-0.440	65%	0.380	0.365-0.394	0.161	0.023
TSA	74%	0.587	0.572-0.602	75%	0.702	0.688-0.717	77%	0.703	0.688-0.718	0.238	0.004
TA	66%	0.429	0.415-0.444	71%	0.476	0.461-0.491	80%	0.487	0.472-0.52	< 0.001a	0.987
VA/TVA	65%	0.470	0.456-0.485	67%	0.472	0.457-0.487	74%	0.542	0.527-0.556	0.016a	0.815

^aP < 0.05 (round 3 vs round 1, uncorrected).

P < 0.0071 vs overall accuracy in round 1 (Bonferroni correction for 7 prespecified comparisons). Data from 1176 polyps assessed by 6 pathologists. κ: < 0.20, slight; 0.21-0.40, fair; 0.41-0.60, moderate; 0.61-0.80, substantial; > 0.80, almost perfect. GCHP: Goblet cell-rich hyperplastic polyp; MVHP: Microvesicular hyperplastic polyp; SSL: Sessile serrated lesion; SSL-D: Sessile serrated lesion with dysplasia; TA: Tubular adenoma; TSA: Traditional serrated adenoma; VA/TVA: Villous adenoma/tubulovillous adenoma.

Interobserver agreement

Interobserver agreement for colorectal polyp diagnoses across the three rounds ranged from fair to moderate, with overall κ values of 0.393 (95%CI: 0.385-0.400) in round 1, 0.441 (95%CI: 0.434-0.449) in round 2, and 0.415 (95%CI: 0.407-0.423) in round 3 (Table 2). Subtype-specific analyses revealed low-fair agreement for the SSLs (κ: 0.301-0.441) and slight agreement for the SSL-D (κ: 0.114-0.132), whereas the highest concordance was observed for the TSAs (κ: 0.587-0.703) (Table 2). The results of the analysis stratified by expertise level revealed that senior pathologists achieved the best agreement across all rounds (polyps overall κ: 0.517-0.441; SSL: 0.601-0.535; SSL-D: 0.105-0.297), whereas the performance of intermediate pathologists for SSL assessment was poor (κ: 0.081-0.035 in rounds 1-3). All pathologists with different levels of experience showed poor-low interobserver agreement for SSL-D in round 1 (κ: 0.084-0.149, Figure 3B).

Intraobserver agreement

The results of the intraobserver agreement among the six pathologists in the first two rounds varied greatly in terms of the diagnosis (Table 3). For overall diagnoses, the senior pathologists exhibited the highest agreement (Doc1: κ = 0.610; Doc2: κ = 0.733), whereas the junior pathologists showed the lowest consistency (Doc5: κ = 0.452; Doc6: κ = 0.432). Regarding the SSL classification, the senior and intermediate pathologists maintained high agreement (Doc1: κ = 0.725; Doc2: κ = 0.742; Doc4: κ = 0.733), whereas the junior pathologists showed moderate agreement (Doc5: κ = 0.517; Doc6: κ = 0.664). However, one intermediate pathologist (Doc3) exhibited low agreement (κ = 0.287). For SSL-D diagnoses, only the senior pathologists (Doc1 and Doc2) achieved moderate agreement (κ = 0.416 and 0.508), whereas the other pathologists displayed poor consistency (κ < 0.4).

Table 3 Analysis of intraobserver agreement (κ) between the first and second reading rounds.

Pathologists	Doc	Overall	SSL	SSL-D
Senior	Doc1	0.610	0.725	0.122
	Doc2	0.733	0.742	0.416
Intermediate	Doc3	0.594	0.287	0.508
	Doc4	0.659	0.733	0.229
Junior	Doc5	0.452	0.517	0.066
	Doc6	0.432	0.664	0.082

Doc: Doctor; SSL: Sessile serrated lesion; SSL-D: Sessile serrated lesion with dysplasia.

Determinants of SSL/SSL-D diagnostic concordance

The univariate analysis revealed that a cross-sectional crypt orientation (consensus vs discordance: 7.4% vs 37.4%, P < 0.001), a serrated region extending to the lower 1/3 of the crypt (96.3% vs 73.7%, P = 0.001), dilation in the lower 1/3 of the crypt (98.1% vs 73.7%, P < 0.001), an abnormal crypt architecture (98.1% vs 70.7%, P < 0.001), > 3 serrated crypts (40.7% vs 22.2%, P = 0.016), and nuclear alterations (7.4% vs 34.3%, P < 0.001) were potential predictors of diagnostic concordance between the SSLs and SSL-D. Age, sex, lesion size, the cell proliferative zone at the basal crypt, and prominent cytoplasmic eosinophilia were not associated with diagnostic concordance. Owing to multicollinearity among the crypt morphological variables (VIF > 5), only the serrated region was retained in the multivariate analysis (Table 4). The multivariate logistic regression analysis revealed that nuclear alterations were the strongest independent predictor of SSL/SSL-D discordance (OR = 10.45; 95%CI: 3.01-36.30; P < 0.001), followed by the cross-sectional crypt orientation (OR = 5.27; 95%CI: 1.61-17.18; P = 0.006). Conversely, the presence of > 3 serrated crypts (OR = 0.37; 95%CI: 0.15-0.92; P = 0.033) and serrated regions extending to the lower 1/3 of the crypt (OR = 0.20; 95%CI: 0.04-0.96; P = 0.044) increased the diagnostic consensus between the SSLs and SSL-D (Table 4).

Table 4 Logistic regression analysis for diagnostic consensus for sessile serrated lesion and sessile serrated lesion with dysplasia, mean ± SD/n (%).

Variable		Consensus group	Discordance group	Univariate analysis	Multivariate analysis
		n = 54	n = 99	P value	P value	OR (95%CI)
Age (years)		53.22 ± 11.50	55.59 ± 13.89	0.274	-	-
Sex	Female	19 (35.2)	38 (38.4)	0.696	-	-
Flat or sessile morphology	Yes	50 (92.6)	95 (96.0)	0.371	-	-
Lesion size1	≤ 0.5 cm	15 (34.1)	32 (38.1)	0.164	-	-
Crypt orientation	Cross-sectional	4 (7.4)	37 (37.4)	< 0.001a	0.006b	5.27 (1.61-17.18)
Serrated region	Lower 1/3 crypt	52 (96.3)	73 (73.7)	0.001a	0.044b	0.20 (0.04-0.96)
Crypt dilation	Lower 1/3 crypt	53 (98.1)	73 (73.7)	< 0.001a	-	-
Proliferative zone	Basal crypt	1 (1.9)	6 (6.1)	0.234	-	-
Abnormal crypt architecture	Yes	53 (98.1)	70 (70.7)	< 0.001a	0.033b	0.37 (0.15-0.92)
Number of serrated crypts	> 3	22 (40.7)	22 (22.2)	0.016a	-	-
Nuclear alteration	Yes	4 (7.4)	34 (34.3)	< 0.001a	< 0.001b	10.45 (3.01-36.30)
Cytoplasmic eosinophilia	Prominent	24 (44.4)	43 (43.4)	0.904	-	-

¹Polyp size data of some cases are missing.

^aP < 0.1.

^bP < 0.05.

OR: Odds ratio.

DISCUSSION

The diagnostic consistency of colorectal serrated lesions, especially SSLs and SSL-D, poses significant challenges in the prevention of CRC. According to our study, the problems associated with the diagnosis of these lesions include morphological complexity, professional experience, and other diagnostic issues.

Serrated lesions composed mainly of SSLs and SSL-D are increasingly recognized as important precursors of CRC via the serrated pathway. They represent a biological entity that is different from the adenoma-carcinoma pathway[14,15]. The serrated pathway has distinct molecular drivers, including the early occurrence of BRAF V600E mutations associated with serrated alterations and the CpG island methylation phenotype associated with aggressive behavior of the tumor[16,17]. As SSLs are closely related to interval CRC[18,19], the precise diagnosis of SSLs is clinically important. Our data revealed lower diagnostic accuracies for the SSLs (62%) and SSL-D (33%) than for the overall polyp. The findings revealed the significance of the diagnostic accuracy of SSL-D compared with that of other polyps after correction for multiple comparisons, with P < 0.001. This analysis revealed that the low diagnostic accuracy of serrated lesions is a major factor for ineffective endoscopic screening, which may lead to missed diagnoses and delayed monitoring.

For both SSL and SSL-D classifications, the difference in performance between senior and junior pathologists was significant. Senior pathologists achieved 51% accuracy for SSL-D (junior pathologists, 17%), whereas the accuracy for SSLs differed by 14% between these groups (69% vs 55%). This interobserver variation likely arises because senior pathologists evaluate a set of features in tissues (for example, asymmetries in SSLs and focal nuclear atypia in SSL-D cases), whereas junior pathologists often rely on a single feature (for example, serrated patterns). Despite an overall increase in accuracy (56%-70%), the performance of junior pathologists plateaued or even declined during SSL/SSL-D recognition across evaluation rounds; this finding highlights the difficulty in identifying these lesions and the need for targeted training[11,20]. The American Society of Gastroenterology and the American College of Gastroenterology have also explicitly included the SSL detection rate in the quality indicators of colonoscopy, which, like the common adenoma, raises the diagnostic requirements for pathologists. Thus, the educational training of pathologists should be repeated, especially for less experienced pathologists, to increase the accuracy and consistency of the diagnosis. A Dutch study revealed that e-learning modules that are mandatory appear to improve the consistency of laboratory diagnostics by 70%[21]. In other words, gaining experience and learning is highly important. However, the differences between pathologists with different levels of experience are not always significant. Compared with that of junior pathologists, the performance of intermediate pathologists on diagnosing SSL cases was not more effective. Even advanced pathologists have less-than-ideal consistency in diagnosing SSL-D cases. Clinical experience alone cannot solve every problem; the results of this study showed that clearer diagnostic criteria and technical support could also prove indispensable.

Morphologically, SSLs appear as flat or slightly elevated lesions, accompanied by serrated glandular structures and basal crypt dilation, which poses diagnostic challenge[22]. The change in diagnostic terminology in the 2019 WHO classification from sessile serrated adenoma to sessile serrated lesion indicates efforts to standardize the criteria[23,24]. Nevertheless, the WHO criteria have not yet been widely adopted, leading to differences in clinical settings. In this study, the diagnostic consistency was still suboptimal, even when the standard criteria were used. The interobserver agreement among gastrointestinal pathologists was fair to moderate for overall polyps (κ = 0.393-0.441). The agreement was particularly poor for the SSLs (κ = 0.409 decreasing to 0.301 across rounds) and SSL-D (κ = 0.114-0.127). Vennelaganti et al[25] also reported diagnostic differences in serrated lesions between pathologists in the United States and Europe, with American pathologists showing higher kappa values for serrated lesions. We observed that SSL-D had the lowest consistency (κ < 0.2), likely due to the unclear boundary of mild atypical hyperplasia. SSLs and SSL-D are easily confused with MVHPs or TSAs. For example, MVHPs with superficial serration are frequently misclassified as SSLs, whereas an SSL-D with focal atypia may be misclassified as a TSA. Additionally, the morphological features of SSLs with gray areas may further increase the complexity of classifying SSLs and SSL-D, and pathologists often hesitate to upgrade their diagnosis without clear atypical hyperplasia, emphasizing the need for clearer diagnostic criteria. Our regression analysis indicated that the use of quantitative criteria, such as the need for > 3 serrated glands to extend to the lower third of the glandular lumen in SSL and SSL-D lesions, can reduce ambiguity (OR = 0.37).

In this study, we analyzed factors that affect diagnostic consistency and found that nuclear alterations and cellular atypia often lead to diagnostic controversies. Although the correlations between these features and the uncertainty and controversy of diagnosis have been mentioned in previous studies[26], they have not yet been systematically addressed. The cross-sectional orientation of crypts contributes to diagnostic controversies, underscoring the importance of proper tissue processing and sampling for achieving an accurate diagnosis. For instance, Tran et al[11] indicated that when samples are not histologically well oriented, such as tangentially or horizontally cut samples, the bases of the crypts are usually not visible. The cross-sectional crypt orientation makes the assessment of architectural features, which are crucial for diagnosing SSLs, challenging. The present findings emphasize the need for comprehensive sampling techniques to ensure that pathologists have sufficient material to make accurate assessments. Notably, the cross-sectional orientation of crypts is a mere technical artifact that relates to tissue processing and sectioning. It does not indicate the incompetence of a pathologist. This finding shows that even the most experienced pathologists can be impeded by suboptimal slide preparation, which needs to be addressed at the laboratory level.

During the third round of evaluation, anonymized clinical data (patient’s age, lesion location, lesion size and endoscopic morphology) were incorporated to better simulate clinical practice. When this additional information was provided, the overall diagnostic accuracy improved from 66% to 72% (P < 0.001), likely because the clinical scenario helped the pathologist narrow a differential diagnosis. Nonetheless, the improvement was not equal: The SSL-D accuracy increased from 33% to 49% (P = 0.007) but remained low, and the interobserver agreement for SSL-D cases remained unchanged (κ: 0.114 in round 1 vs 0.132 in round 3). The divergent trajectories in the SSL and SSL-D classifications likely reflect a shift in reclassification when additional clinical information is considered. Clinically, larger lesion sizes may have prompted pathologists to upgrade some SSL diagnoses to SSL-D, thereby increasing SSL-D accuracy. Moreover, other SSLs may be diagnosed as other types due to factors such as the lesion location, significantly decreasing SSL accuracy. An ambiguous diagnosis of serrated lesions cannot be solved fundamentally by this trade-off based on clinical information. The diagnosis of SSL-D is based on subtle cellular abnormalities that cannot be elucidated from the clinical data. Based on these findings, the essential difficulty in diagnosing SSL-D cases is the recognition of histological features and not clinical information.

The moderate diagnostic consistency (κ = 0.409) of SSL cases addressed here also highlights the current histological limitations. Although auxiliary biomarkers such as AGRN and CD133 improve the SSL diagnosis via immunohistochemical staining[27,28], their clinical application has thus far been limited because of the large volume of colonoscopy biopsies and high reagent costs. In contrast, artificial intelligence (AI)-assisted diagnostic tools can perform scalable analyses of regular HE-stained sections in an efficient and cost-effective manner. This finding is particularly important, as an increasing number of colorectal polyp biopsies are performed. Digital pathology and AI techniques have great potential for diagnosing SSLs/SSL-D. AI algorithms trained on annotated SSL and SSL-D images can quantify the number of serrated crypts, degree of crypt expansion, and areas of dysplasia features in which pathologists often differ in terms of the criteria for actual interpretation. AI can play the role of an alarm system for an inconsistent diagnosis to help reduce the difference in the prediction. AI, on the one hand, provides the distribution of predicted probabilities and, on the other hand, prioritizes the recall of SSL-D cases with extremely low sensitivity through outlier detection. As many as 80%-95% of the SSLs were correctly identified from other colorectal polyps by some prototypical models[29,30]. For SSL-D cases, the AI algorithms can assist in generating a map of the distribution of nuclear heterogeneity, thereby minimizing the reliance on subjective grading criteria. Moreover, the construction of a consensus validated SSL/SSL-D shared digital library through collaboration with multiple institutions will promote continuous learning and calibration across centers, which will help resolve the diagnostic differences and low performance of our multicenter queue.

Although this study clarifies the difficulties of SSL/SSL-D diagnoses, the retrospective design limits the applicability of the results in practice. First, the problem with using expert consensus as the reference standard is that the expert agreement can be wrong. Future studies should incorporate molecular annotations (e.g., BRAF mutation and MLH1 methylation) for discrepant cases of high clinical relevance, such as those suspected of being SSL-D. Furthermore, latent class models can be used to estimate the true diagnostic performance even in the absence of a perfect gold standard. Second, the retrospective design limits the generalizability of the findings in clinical settings; prospective trials are needed to validate the prognostic value of the results. Third, as a pathology-focused study, we did not collect detailed information on endoscopic factors such as the equipment type, endoscopist experience level, or bowel preparation quality, all of which could influence lesion detection and sample quality. Furthermore, the staged diagnostic process was performed on archived samples and did not influence patient care; prospective studies are needed to assess the clinical utility of improved diagnostic consistency. Additionally, our three-round design cannot fully separate the effect of repeated exposure from the added clinical information; a crossover design is needed to address this issue in future studies. Other residual confounding factors that were not controlled for in this analysis included intercenter variability in tissue processing and sectioning techniques, the prior case exposure of individual pathologists, and the potential training effect of panel discussions between rounds.

In addition to the methodological limitations outlined above, the diagnostic classification in this study was based on the WHO 5^th edition classification. The basic morphological criteria for SSL/SSL-D cases have not changed in the newly released 6^th edition, and our research results do not lessen their effectiveness with the publication of the new classification. In contrast, they may be valuable supplements to the implementation of the new WHO classification criteria. Through our observations of ongoing diagnostic differences among experts, we indicate that merely updating the classification system without addressing implementation issues is futile. Focusing on learning and unifying morphological diagnostic criteria and enforcing the orientation of the slice is required. In future studies, standardization of the entire process of endoscopic and pathological testing should be explored to further improve the detection and diagnosis rates of serrated lesions.

CONCLUSION

The classification criteria and the histological complexity of SSL and SSL-D cases reduce the prevention of CRC. A multipronged approach is needed to reduce this gap: Quantitative thresholds should inform diagnostic criteria, the subspecialty training of pathologists should be expanded, and the use of AI algorithms should offset human variability. By overcoming these challenges, the detection and diagnosis rates of serrated lesions can be improved, leading to a reduced global burden of interval CRC.

ACKNOWLEDGEMENTS

The authors would like to thank all study members and the participants.