Agreement Between WebCeph and Manual Cephalometric Analysis: Cleft and Non‐Cleft Patients Comparison: A Retrospective Method‐Comparison Study Authors: Aylar Afshari, Zahra Mohandes, Shabnam Ajami Published: 2026 ABSTRACT ABSTRACTBackground and AimsCephalometric analysis is essential in orthodontic diagnosis, but landmark identification is particularly challenging in cleft lip and palate (CLP). Manual tracing remains, however, time‐consuming and operator‐dependent, the gold standard. AI‐based systems such as WebCeph may improve efficiency, though agreement with manual tracing remains unclear. This study evaluated agreement between WebCeph and manual tracing in complete unilateral CLP (CUCLP) patients, comparing discrepancies with non‐cleft controls.MethodsThis retrospective method‐comparison study included lateral cephalograms of 45 children with repaired CUCLP and 30 age‐matched controls (8–12 years). Seven angular measurements based on 14 landmarks were analyzed using manual tracing and WebCeph. Agreement was assessed using the Intraclass correlation coefficient (ICC), group differences were tested with parametric or non‐parametric tests, measurement error was calculated using Dahlberg's formula.ResultsOf 75 cephalograms (mean age 8.85 ± 1.43 years), WebCeph failed in 17 cases (22.6%), with similar failure rates in cleft (22.2%) and non‐cleft (23.3%) groups, leaving 58 radiographs for analysis. ICC ranged from 0.332 to 0.845: good for SNB, and ANB; moderate for SNA and U1–SN; and poor for gonial angle, IMPA, and nasolabial angle. Significant differences occurred for gonial angle, nasolabial angle, U1–SN, and IMPA in cleft patients and for SNA in controls (p < 0.05), with no significant between‐group difference in disagreement magnitude (p > 0.05).ConclusionWebCeph showed variable agreement with manual tracing, from moderate to good for skeletal measurements (SNA, SNB, ANB) to poor for gonial angle, IMPA, and nasolabial angle. Similar discrepancy patterns, failure rates, and disagreement magnitudes across cleft and non‐cleft groups suggest algorithmic limitations as a contributing factor, though a meaningful influence of cleft‐related anatomy cannot be excluded given the sample size. AI‐based cephalometry should be used as an adjunct rather than a replacement for manual tracing in clinical decision‐making. Introduction Cleft lip and palate (CLP) is a common craniofacial anomaly that requires long‐term multidisciplinary management [1, 2]. Orthodontic treatment planning in these patients depends heavily on cephalometric analysis because craniofacial growth disturbances and skeletal discrepancies are common [3]. Accurate identification of cephalometric landmarks is therefore essential for diagnosis, treatment planning, and outcome assessment [4]. Conventional manual tracing remains the standard method for cephalometric analysis [5]. However, the process is time‐consuming and influenced by examiner experience and landmark identification variability [6]. These challenges become more pronounced in patients with complete unilateral cleft lip and palate (CUCLP), in whom altered anatomy, asymmetry, overlapping structures, and reduced radiographic clarity may complicate landmark localization [7]. Artificial intelligence (AI)‐based cephalometric software has been introduced to improve workflow efficiency and reduce operator dependency. WebCeph (AssembleCircle Corp, Seoul, Korea) is a web‐based orthodontic platform capable of automated landmark identification and cephalometric measurement generation [8, 9, 10]. Previous studies evaluating AI‐assisted cephalometric analysis have reported inconsistent findings. Some investigations demonstrated acceptable agreement between AI‐generated and manual measurements, whereas others identified clinically relevant discrepancies, particularly for dental and soft tissue parameters [6, 10, 11, 12, 13]. Landmark variability remains a major limitation regardless of tracing method, with reproducibility affected by radiographic clarity and structural superimposition in both manual and AI‐assisted analyses [14, 15, 16]. Although several studies have evaluated AI‐based cephalometric software, evidence regarding its performance in cleft populations remains limited. In addition, few studies have compared disagreement patterns between cleft and non‐cleft individuals or assessed software operational failures [17, 18, 19]. Understanding both measurement agreement and practical applicability is important before AI‐generated cephalometric analysis can be integrated into routine orthodontic care. Although several studies have evaluated AI‐assisted cephalometric analysis in general orthodontic populations, evidence regarding its performance in patients with CUCLP remains limited. The complex craniofacial anatomy and altered anatomical landmarks characteristic of this population may reduce the agreement of automated landmark identification and cephalometric measurements [15]. Consequently, the reliability and practical utility of AI‐assisted cephalometric analysis in patients with CUCLP remain uncertain. The aim of this study was to evaluate the clinical applicability of WebCeph by assessing its agreement with manual cephalometric analysis and its software processing success rate in patients with CUCLP compared with non‐cleft controls. Materials and Methods Ethical Approval and Study Design This retrospective method‐comparison study evaluated the agreement and operational performance between manual cephalometric tracing and AI‐based software. Ethical approval was obtained (IR.SUMS.DENTAL.REC.1404.034), and the study was conducted in accordance with the Declaration of Helsinki and reported following the STARD guideline. Written informed consent was obtained from the legal guardians of all included minors. Patient data were retrieved from the Cleft Lip and Palate Registry System at the Cleft Lip and Palate Center. The study included patients with unilateral CUCLP and an age‐matched non‐cleft control group. The control group consisted of skeletal Class I subjects (ANB ≈ 2° ± 2°) with no history of craniofacial anomalies or orthodontic treatment, selected from routine clinical records. Age matching was performed at the group level rather than through individual one‐to‐one pairing; controls were consecutively selected within a comparable age range to the CUCLP group (7–12 years) to achieve comparable group‐level age distributions. Inclusion of a non‐cleft group allowed assessment of whether cleft‐related anatomical complexity influences discrepancies between manual and AI‐based cephalometric measurements. All eligible cephalograms were consecutively included. Inclusion criteria comprised non‐syndromic CUCLP patients after cleft repair and non‐cleft skeletal Class I subjects. Exclusion criteria included bilateral cleft lip and palate, craniofacial trauma, systemic diseases, and unrelated surgical interventions [20]. All radiographs were obtained as part of routine clinical assessment and not for research purposes, in accordance with the ALARA principle. For CUCLP patients, additional clinical information was recorded from archived records, including the timing and type of previous lip and palate repair, alveolar bone graft status, and whether orthodontic treatment was active at the time of radiographic acquisition, as these factors may influence anatomical landmark visibility on lateral cephalograms. Cephalometric Analysis All lateral cephalograms were obtained at a single radiographic center using a standardized cephalostat (Orthopantomograph OP200, USA), with patients positioned in natural head position and teeth in centric occlusion. Images were acquired at a maximum voltage of 85 kV [21]. Each cephalogram served as its own control, with manual and AI measurements from the same radiograph, allowing for within‐subject comparison without randomization. Fourteen anatomical landmarks (11 hard tissue, 2 dentals, and 1 soft tissue) were identified, and seven angular cephalometric measurements were analyzed (Table 1 and Figure 1) [22]. Table 1Hard tissue, soft tissue, and dental cephalometric measurements and definitions.Cephalometric measurementsDefinitionHard tissueSNASagittal position of the maxilla relative to the anterior cranial base plane S‐NSNBSagittal position of the mandible related to the anterior cranial base plane S‐NANBReciprocal relation between the upper and lower jaw on the sagittal planeGonial angleThe angle between these three points: articular, gonion, and gnathionDental measurementsU1SNUpper incisor inclination relative to the SN planeIMPALower incisor inclination relative to the mandibular plane (incisor mandibular plane angle)Soft tissueNasolabial angleThe angle between these three points: anterior nasalis, subnasalis, and labiale superiorisJohn Wiley & Sons, Ltd. Hard tissue, soft tissue, and dental cephalometric measurements and definitions. Anatomical landmarks used for manual lateral cephalometric tracing, illustrating the reference points required for Steiner's analysis on a representative lateral cephalogram. All eligible cases within the predefined study period were included consecutively. Although a formal a priori sample size calculation was not performed, a post‐hoc precision assessment was conducted based on the achieved sample size (n = 58). The 95% confidence interval widths for the ICC estimates ranged from 0.16 (ANB) to 0.80 (IMPA), reflecting substantially greater precision for parameters with higher agreement (SNB, ANB) and wider, less precise intervals for parameters with poorer agreement (IMPA, nasolabial angle). This indicates the sample size was sufficient to detect meaningful differences in agreement for skeletal measurements, but gave less precise estimates for the gonial angle, IMPA, and nasolabial angle, so the poor‐agreement findings for these parameters should be interpreted with caution. Manual tracing was performed on acetate overlays placed directly on printed lateral cephalograms, using a 0.5 mm mechanical pencil and a standard cephalometric ruler. All printed radiographs were produced at a 1:1 scale from digital images exported from the radiographic system, and no additional magnification correction or digital calibration was required. Landmark identification and angular measurements were performed by a trained examiner according to Steiner's cephalometric analysis, following a standardized and pre‐defined tracing protocol (Figure 2A). The examiner was unaware of the AI‐generated measurements throughout the process; however, complete blinding to group allocation was not feasible, as cleft‐related anatomical features were inherently visible on the lateral cephalograms, which may have introduced a degree of observer bias. All completed tracings were subsequently reviewed by an experienced orthodontist, who verified landmark placement and measurement values to reach a consensus. Because the second examiner reviewed for consensus purposes rather than conducting an independent set of measurements, inter‐examiner reliability could not be formally assessed and represents an acknowledged limitation of this study. One‐third of cases were randomly selected for repeat tracing by the primary examiner to assess intra‐examiner reliability, calculated using Dahlberg's formula [23]. The images were used at their original scale, and no additional digital calibration was required. One‐third of cases were randomly selected for repeat analysis to assess intra‐examiner reliability. Inter‐examiner reliability was not evaluated, as consensus‐based tracing was used. Manual tracing was considered the reference method, although inherent variability in landmark identification is acknowledged. Representative lateral cephalometric tracing of a single CUCLP patient, comparing (a) manual tracing performed by the primary examiner and (b) automated landmark identification and tracing generated by WebCeph. Traced lines and points indicate the anatomical landmarks and reference planes used for Steiner's analysis. All identifying facial features have been removed/obscured. AI‐based cephalometric analysis was performed using WebCeph (web version; AssembleCircle Corp., Seoul, South Korea), which automatically identified anatomical landmarks and generated measurements based on Steiner's analysis (Figure 2B) [24]. All AI‐assisted analyses were performed between September and October 2025. The AI‐generated landmarks were accepted without manual adjustment, allowing evaluation of the software's fully automated performance. Cases in which the software failed to generate usable outputs were recorded and excluded from agreement analysis but reported separately as part of operational performance. Because AI failures were considered part of the operational performance evaluation, these cases were retained for failure‐rate analysis but excluded from agreement analyses. Statistical Analysis Statistical analysis was performed using SPSS version 22. Agreement between manual and AI‐based measurements was assessed using intraclass correlation coefficients (ICC) based on a two‐way mixed‐effects model with absolute agreement. Bland–Altman analysis was used to evaluate bias and 95% limits of agreement for each cephalometric variable. Intra‐examiner reliability was assessed separately from the method‐comparison analysis, using the repeat tracings performed by the primary examiner on one‐third of the sample. For each of the seven cephalometric variables, agreement between the first and second tracing was quantified using Dahlberg's formula (√(Σd 2/2n) [23], where d is the difference between the two measurements for a given case and n is the number of cases) and by the Intraclass correlation coefficient (ICC), calculated using the same two‐way mixed‐effects, absolute‐agreement model applied to the method‐comparison ICCs above. Proportional bias for each variable was assessed by regressing the difference between methods against their mean, and heteroscedasticity was evaluated by examining whether the magnitude of these differences varied systematically across the measurement range. Differences between methods were explored using paired t‐tests or Wilcoxon signed‐rank tests, depending on data distribution; however, these were considered exploratory and not true agreement analyses. Between‐group comparisons (cleft vs. non‐cleft) were performed using independent t‐tests or Mann–Whitney U tests. Normality of the paired differences was assessed using the Shapiro–Wilk test. For angles with normally distributed differences, the mean difference and its t‐distribution‐based 95% confidence interval are reported; for non‐normally distributed differences (IMPA in the CUCLP group and the gonial angle in the control group; Shapiro–Wilk p < 0.05), the median difference and interquartile range (IQR) are reported, with the 95% confidence interval for the median obtained by bootstrap resampling (10,000 replicates). Given the exploratory nature of the paired comparisons across seven cephalometric angles, no correction for multiple comparisons (e.g., Bonferroni) was applied; p values for these comparisons should therefore be interpreted as descriptive rather than confirmatory, and primary inference was based on ICC estimates and their 95% confidence intervals rather than on hypothesis testing. All statistical tests were two‐sided, and a p value < 0.05 was considered statistically significant a priori. Group differences in WebCeph processing failure rates between cleft and non‐cleft patients were compared using Fisher's exact test. All statistical analyses were performed using IBM SPSS Statistics, version 22.0 (IBM Corp., Armonk, NY, USA). The primary analysis, agreement assessment via ICC and Bland–Altman methods, was prespecified. All paired and between‐group comparisons of individual angle measurements were exploratory and were not part of the prespecified analysis plan. Results Demographic and baseline characteristics of both groups are summarized in Table 2. The groups did not differ significantly in age (p = 0.92) or sex distribution (p = 0.11). A total of 75 lateral cephalometric radiographs were analyzed, with a female‐to‐male ratio of 40/35. The patients' ages at the time of imaging ranged from 7 to 12 years (mean = 8.85 ± 1.43). Participants were divided into two groups: 45 patients with unilateral complete cleft lip and palate (CUCLP) and 30 non‐cleft controls with skeletal Class I relationships and no orofacial anomalies. The participant selection process and reasons for exclusion are summarized in the STARD flow diagram (Figure 3). Statistical analyses and reporting were conducted in accordance with recommendations for reporting statistics in clinical research. Table 2Demographic and baseline characteristics of the CUCLP and non‐cleft control groups.CharacteristicCUCLPa (n b = 35)Control (n b = 23)Test p valueAge (years), mean ± SD9.06 ± 1.688.96 ± 1.22Mann–Whitney U test (age distributions were non‐normal by Shapiro–Wilk)0.92Age range (years)7–128–12——Median age (years)9.09.0——Sex n (%)Female14 (40.0%)15 (65.2%) χ 2 test0.11Male21 (60.0%)8 (34.8%) a CUCLP = complete unilateral cleft lip and palate. b n = number.John Wiley & Sons, Ltd. Demographic and baseline characteristics of the CUCLP and non‐cleft control groups. CUCLP = complete unilateral cleft lip and palate. n = number. Participant flowchart illustrating the selection and exclusion process. Of 75 eligible radiographs (45 complete unilateral cleft lip and palate (CUCLP), 30 non‐cleft), 17 were excluded due to WebCeph processing failure, leaving 58 radiographs available for agreement analysis (35 CUCLP, 23 non‐cleft). Among the 35 patients with CUCLP included in the final analysis, primary lip repair was performed at a mean age of 4.18 ± 1.69 months (range: 0.5–6 months). Palate repair was completed at a mean age of 13.54 ± 6.95 months (range: 9–48) using VY pushback (16, 45.7%), Modified Furlow palatoplasty (10, 28.6%), and Sommerlad palatoplasty (9, 25.7%) techniques. Secondary alveolar bone grafting had been performed in 10 patients (28.57%). At the time of radiographic acquisition, no patients had active orthodontic treatment. Agreement between manual tracing and WebCeph measurements varied considerably across parameters. Single‐measure Intraclass correlation coefficients (ICC) ranged from 0.332 to 0.845 (Table 3). Good agreement was observed for SNB and ANB (ICC 0.838 and 0.845), moderate agreement for SNA and U1–SN (ICC 0.745 and 0.738), and poor agreement for the gonial angle, IMPA, and nasolabial angle (ICC 0.357, 0.332, and 0.368, respectively). All ICC estimates were statistically significant (p < 0.001), though the strength of these associations remained clinically limited for the gonial angle, IMPA, and nasolabial angle. Table 3The agreement between manual cephalometric tracing and WebCeph (AssembleCircle Corp, Seoul, Korea) digital analysis for all angles using the Intraclass Correlation Coefficient (ICC) with a two‐way mixed‐effects model and absolute agreement definition.Angles (n a)ICC value95% confidence intervalInterpretation p valueLower boundUpper boundSNA (58)0.7450.6000.840Moderate< 0.001SNB (58)0.8380.7300.900Good< 0.001ANB (58)0.8450.7500.910Good< 0.001Gonial (58)0.3570.0700.580Poor< 0.001U1SN (58)0.7380.4700.860Moderate< 0.001IMPA (58)0.332−0.1000.670Poor< 0.001Nasolabial (58)0.368−0.0800.720Poor< 0.001 Note: The ICC p value reflects only whether the coefficient differs significantly from zero. The strength of agreement is determined by the magnitude of the ICC rather than its statistical significance; agreement was interpreted according to Koo and Li [25]: ICC < 0.50 = poor, 0.50–0.75 = moderate, 0.75–0.90 = good, and > 0.90 = excellent; for IMPA and the nasolabial angle, the 95% CI for the ICC includes zero despite p < 0.001; this reflects the fact that the significance test and 95% CI for the absolute‐agreement ICC are derived from different F‐ratios and is not indicative of a data error. a n = number.John Wiley & Sons, Ltd. The agreement between manual cephalometric tracing and WebCeph (AssembleCircle Corp, Seoul, Korea) digital analysis for all angles using the Intraclass Correlation Coefficient (ICC) with a two‐way mixed‐effects model and absolute agreement definition. Note: The ICC p value reflects only whether the coefficient differs significantly from zero. The strength of agreement is determined by the magnitude of the ICC rather than its statistical significance; agreement was interpreted according to Koo and Li [25]: ICC < 0.50 = poor, 0.50–0.75 = moderate, 0.75–0.90 = good, and > 0.90 = excellent; for IMPA and the nasolabial angle, the 95% CI for the ICC includes zero despite p < 0.001; this reflects the fact that the significance test and 95% CI for the absolute‐agreement ICC are derived from different F‐ratios and is not indicative of a data error. n = number. Intra‐examiner reliability, assessed by repeat tracing of one‐third of the sample by the primary examiner, was excellent across all seven cephalometric variables (ICC 0.976–0.996), with Dahlberg's measurement error ranging from 0.27° to 0.73° (Table 4). Notably, intra‐examiner reliability remained excellent for the gonial angle, IMPA, and nasolabial angle, the three variables that showed the poorest agreement between manual tracing and WebCeph, indicating that this poor manual‐AI agreement was not attributable to unreliable manual measurement. Table 4Intra‐examiner reliability of manual cephalometric measurements (repeat tracing, n = 19).AnglesDahlberg's error (°)ICC estimates p value95% confidence interval for ICCICC (A,1)Lower boundUpper boundSNA0.270.99> 0.990.996< 0.001SNB0.520.940.990.977< 0.001ANB0.490.930.990.976< 0.001Gonial0.570.94> 0.990.990< 0.001U1SN0.640.99> 0.990.995< 0.001IMPA0.590.97> 0.990.990< 0.001Nasolabial0.730.99> 0.990.996< 0.001 Note: Two‐way mixed‐effects model, absolute agreement, single measures (same model used for the method‐comparison ICCs in Table 3). Dahlberg's error calculated as √(Σd 2/2n), where d is the difference between the two tracings for a given case [23].Abbreviation: ICC = intraclass correlation coefficient.John Wiley & Sons, Ltd. Intra‐examiner reliability of manual cephalometric measurements (repeat tracing, n = 19). Note: Two‐way mixed‐effects model, absolute agreement, single measures (same model used for the method‐comparison ICCs in Table 3). Dahlberg's error calculated as √(Σd 2/2n), where d is the difference between the two tracings for a given case [23]. Abbreviation: ICC = intraclass correlation coefficient. During AI processing, WebCeph failed to generate usable outputs in 17 cases (17/75, 22.6% = 10 CUCLP and 7 controls) (Figure 3). In none of these cases did the software provide an error message or specify a reason for failure; the platform simply did not produce a measurable output. As all radiographs were obtained using a standardized cephalostat under identical acquisition parameters and were of sufficient diagnostic quality for manual tracing, gross image acquisition problems are an unlikely explanation for the failures. However, because WebCeph did not provide error messages or failure codes, the precise cause could not be determined, and subtler image‐ or patient‐related contributors, such as anatomical variation, landmark visibility, or case‐specific artifacts, cannot be excluded. These failures are therefore best characterized as unresolved technical processing limitations of the WebCeph algorithm, rather than definitively attributed to the algorithm alone. To formally compare failure rates between groups, Fisher's exact test was applied. The failure rate in the CUCLP group (10/45, 22.2%) did not differ significantly from that in the non‐cleft control group (7/30, 23.3%) (p > 0.99, odds ratio = 0.94), confirming that processing failure occurred at virtually identical rates regardless of group allocation. These 17 cases were excluded from all subsequent agreement analyses, leaving 58 radiographs available for final comparison (35 CUCLP, 23 non‐clefts). In CUCLP patients, statistically significant differences between manual and WebCeph measurements were observed for the gonial angle, nasolabial angle, U1–SN, and IMPA (p < 0.05), whereas no significant differences were found for other parameters (p > 0.05) (Table 5). In the non‐cleft group, similar patterns were observed, except for SNA, which showed a significant difference between methods (p = 0.04) (Table 6). Overall, the direction and magnitude of discrepancies varied across measurements and were not consistent between groups. Table 5Comparison of angular measurements between manual and WebCeph methods in CUCLP patients (n = 35).AngleMean ± SDa p valueManual methodWebCeph methodSNA76.97 ± 4.0077.66 ± 2.910.22SNB74.34 ± 3.7975.15 ± 3.810.05ANB2.65 ± 4.562.51 ± 3.390.66Gonial126.82 ± 4.97129.71 ± 6.15< 0.001IMPA94.55 ± 6.8385.27 ± 7.36< 0.001U1SN92.91 ± 7.9089.23 ± 8.790.005Nasolabial118.98 ± 13.1299.03 ± 12.99< 0.001 a Standard deviation.John Wiley & Sons, Ltd. Comparison of angular measurements between manual and WebCeph methods in CUCLP patients (n = 35). Standard deviation. Table 6Comparison of angular measurements between manual and WebCeph methods in non‐cleft patients (n = 23).AngleMean ± SDa p valueManual methodWebCeph methodSNA80.91 ± 3.8481.92 ± 3.860.04SNB77.65 ± 3.7778.22 ± 3.920.20ANB3.08 ± 1.223.70 ± 1.670.05Gonial123.97 ± 3.88128.91 ± 6.23< 0.001IMPA97.73 ± 5.6588.03 ± 4.85< 0.001U1SN104.69 ± 9.58100.52 ± 6.34< 0.001Nasolabial125.28 ± 11.02105.74 ± 15.56< 0.001 a Standard deviation.John Wiley & Sons, Ltd. Comparison of angular measurements between manual and WebCeph methods in non‐cleft patients (n = 23). Standard deviation. To further assess group effects, the differences between manual and AI measurements were compared between CUCLP and control groups (Table 7). No statistically significant differences were found between groups for any cephalometric parameter (Table 7). However, the absence of statistical significance should not be interpreted as evidence of equivalence, particularly given the modest sample sizes in each group. Effect estimates with 95% confidence intervals are reported in Table 7 to allow readers to assess the plausible range of between‐group differences. Although point estimates were small for most parameters, the confidence intervals were wide, meaning a clinically meaningful influence of cleft‐related anatomy on measurement disagreement cannot be excluded on the basis of these data alone. Bland–Altman analysis demonstrated variable agreement across parameters. SNA, SNB, and ANB showed relatively narrower limits of agreement, whereas wider variability was observed for gonial angle, U1–SN, IMPA, and nasolabial angle (Table 8). The largest disagreement was observed for the nasolabial angle (bias = −19.79°; 95% limits of agreement: −38.07° to −1.50°), followed by IMPA (bias = −9.45°; −19.82° to 0.92°). These results indicate limited clinical interchangeability between manual and AI‐based measurements, particularly for dental and soft tissue parameters. Bland–Altman plots for IMPA, nasolabial angle, and gonial angle are shown in Figure 4, while plots for SNA, SNB, ANB, and U1–SN are provided in the Supporting Information S1: Figures S1–S4). Table 7Results of differences between the angle values assessed manually and via WebCeph in CUCLP patients versus the control group.Diff.a angleMean ± SDb Mean difference (95% CI) p valuePatients with cleft (n = 35)Control group (n = 23)SNA0.69 ± 3.301.01 ± 2.18−0.32 (−1.73 to 1.09)0.68SNB0.81 ± 2.390.56 ± 2.090.25 (−0.93 to 1.43)0.69ANB−0.15 ± 1.980.61 ± 1.48−0.76 (−1.68 to 0.15)0.12Gonial2.88 ± 4.85Median 2.88 (IQR 0.60 to 8.49)−2.05 (−5.06 to 0.96)0.25IMPAMedian −8.58 (IQR −10.72 to −5.07)−9.70 ± 4.940.42 (−2.28 to 3.12)0.87U1SN−3.68 ± 7.24−4.17 ± 5.160.49 (−2.64 to 3.62)0.76Nasolabial−19.94 ± 8.69−19.54 ± 10.41−0.40 (−4.98 to 4.18)0.32 a Diff. = WebCeph—manual method. b Standard deviation, n = number of patients; the p value of no angle showed significant differences (p value > 0.05); in patients with cleft, for IMPA and in the control group for gonial angle, the paired difference was non‐normally distributed (Shapiro–Wilk p = 0.002); median (IQR) and a bootstrap 95% CI are reported. All other angles: mean difference with t‐based 95% CI.John Wiley & Sons, Ltd. Results of differences between the angle values assessed manually and via WebCeph in CUCLP patients versus the control group. Diff. = WebCeph—manual method. Standard deviation, n = number of patients; the p value of no angle showed significant differences (p value > 0.05); in patients with cleft, for IMPA and in the control group for gonial angle, the paired difference was non‐normally distributed (Shapiro–Wilk p = 0.002); median (IQR) and a bootstrap 95% CI are reported. All other angles: mean difference with t‐based 95% CI. Table 8Bland–Altman summary of bias and 95% limits of agreement between manual tracing and WebCeph for all cephalometric measurements (n = 58).MeasurementBiasa Lower 95% LoAb Upper 95% LoASNA0.82−4.856.49SNB0.72−3.725.15ANB0.16−3.433.74Gonial3.7−7.7615.15U1–SN−3.88−16.538.78IMPA−9.45−19.820.92Nasolabial−19.79−38.07−1.5 a Bias = WebCeph—manual. b LoA = limits of agreement.John Wiley & Sons, Ltd. Bland–Altman summary of bias and 95% limits of agreement between manual tracing and WebCeph for all cephalometric measurements (n = 58). Bias = WebCeph—manual. LoA = limits of agreement. Bland–Altman plots for IMPA, nasolabial angle, and gonial angle comparing manual tracing and WebCeph (n = 58). The solid horizontal line represents the mean difference (bias), and the dashed lines indicate the 95% limits of agreement. The x‐axis represents the mean of the manual and WebCeph measurements for each parameter (in degrees), and the y‐axis represents the difference between the two methods (WebCeph—manual, in degrees). Proportional bias was identified for ANB (slope = 0.24, p = 0.001) and the gonial angle (slope = −0.35, p = 0.03), indicating that the magnitude of disagreement for these two parameters varied systematically with the measured value rather than remaining constant across the range. Heteroscedasticity was additionally identified for ANB (p = 0.03). For the remaining five parameters (SNA, SNB, U1–SN, IMPA, and nasolabial angle), no significant proportional bias or heteroscedasticity was detected, supporting the assumption of constant bias and limits of agreement across the measurement range for these variables. Discussion This study evaluated the agreement between WebCeph and manual cephalometric tracing in patients with complete unilateral cleft lip and palate (CUCLP) and compared patterns of disagreement with those in non‐cleft controls. Agreement between methods varied by measurement type: skeletal parameters (SNA, SNB, ANB) showed moderate‐to‐good agreement, whereas the gonial angle, IMPA, and nasolabial angle showed poor agreement [26, 27]. In addition, WebCeph failed to generate usable outputs in 22.6% of cases (17/75), representing an important operational limitation. The similarity of disagreement patterns between cleft and non‐cleft groups suggests that algorithmic limitations may represent a primary contributor to the observed discrepancies; however, this interpretation should be approached with caution. The absence of a statistically significant between‐group difference does not confirm equivalence, and the wide confidence intervals around effect estimates indicate that the present study was not sufficiently powered to exclude a meaningful influence of cleft‐related anatomy on measurement disagreement. A contribution of cleft‐specific anatomical complexity to the observed discrepancies therefore cannot be ruled out [23, 27]. Although craniofacial asymmetry and altered skeletal morphology in CUCLP patients may complicate landmark identification, comparable disagreement in controls suggests that current AI‐based systems remain affected by broader limitations in automated landmark detection [27]. The ICC values confirm that agreement between manual and AI‐generated measurements is parameter‐dependent, with WebCeph reasonably reliable for skeletal relationships but not interchangeable with manual tracing for the gonial angle, IMPA, and nasolabial angle. Even where mean differences were small, the wide Bland‐Altman limits of agreement indicate the two methods cannot be considered clinically interchangeable. Applying a commonly cited clinically acceptable error threshold of ±2° for cephalometric angular measurements [28], the mean bias for SNA, SNB, and ANB fell within this threshold, whereas bias for the gonial angle, U1–SN, IMPA, and the nasolabial angle substantially exceeded it. Notably, even for parameters with mean bias within this threshold, the corresponding Bland–Altman limits of agreement were considerably wider than ±2° (Table 8), indicating that low average bias did not translate into narrow individual‐level disagreement. This distinction, between small average bias and wide case‐by‐case variability, has direct clinical relevance, as low mean bias alone does not guarantee that any single patient's measurement will be clinically acceptable. This threshold, commonly applied to skeletal angular measurements, may not be equally appropriate for dental and soft‐tissue parameters, which are subject to greater anatomical and tracing variability. Several previous studies evaluating AI‐assisted cephalometric analysis have reported mixed findings. Some investigations demonstrated acceptable agreement between AI‐generated and manual measurements, whereas others identified significant discrepancies, particularly for dental and soft tissue variables [11, 27, 29, 30, 31, 32]. Differences among studies may reflect variations in sample characteristics, landmark definitions, imaging quality, statistical methodology, and clinician supervision during AI analysis [17, 33]. Another factor that may have contributed to the observed variability is the use of Steiner cephalometric analysis. Although Steiner analysis is widely used in orthodontic practice and is incorporated into the WebCeph platform, enabling a direct comparison between manual and AI‐generated measurements, it was not specifically developed for patients with CLP. Landmarks such as Point A and measurements involving the maxillary dentoalveolar region may be more difficult to identify because of cleft‐related anatomical alterations. Therefore, some of the observed disagreement may reflect the challenges of applying a conventional cephalometric analysis to a cleft population in addition to the limitations of the AI algorithm itself. Future studies should investigate AI performance using cephalometric analyses specifically developed for cleft patients, such as the Delaire analysis. Because manual tracing and WebCeph analysis were performed on the same radiographic image for each patient, clinical factors such as bone graft status and the timing or type of prior cleft repair were identical across both methods for any given case; these factors are therefore unlikely to explain the discrepancy between manual and AI‐based measurements, even though they may still contribute to the general difficulty of landmark identification in this population. In the present study, gonial angle, nasolabial angle, U1–SN, and IMPA demonstrated significant differences between methods in cleft patients. Similar discrepancies were also observed in non‐cleft individuals. Dental and soft tissue measurements appear particularly vulnerable to landmark identification variability because they depend on the precise localization of incisal apices, mandibular contours, and soft tissue boundaries [31]. IMPA showed considerable disagreement between methods. Accurate identification of the mandibular incisor apex and the mandibular plane remains challenging because of variability in incisor angulation and mandibular contour definition, particularly in patients with altered dentoalveolar anatomy. Similar limitations in AI‐assisted dental landmark measurements have been reported in previous studies [25, 33, 34]. The Bland–Altman analysis further demonstrated a substantial systematic bias (−19.79°) and very wide limits of agreement (−38.07° to −1.50°) for the nasolabial angle. From a clinical perspective, discrepancies of this magnitude exceed what would generally be considered acceptable for orthodontic diagnosis and treatment planning. Therefore, nasolabial angle values generated by WebCeph should not be regarded as clinically interchangeable with manual measurements in patients with CUCLP, and its use in clinical decision‐making should be approached with caution. IMPA and U1‐SN also demonstrated meaningful disagreement, with a moderate systematic bias for IMPA (−9.45°) and wider limits of agreement for U1‐SN. Misestimation of these parameters may affect the evaluation of incisor inclination and soft‐tissue profile assessment, particularly in cleft patients undergoing dentoskeletal correction [11, 27, 31, 32]. The operational failure rate of WebCeph, affecting nearly one‐quarter of radiographs in both cleft and non‐cleft patients, is itself a clinically important finding, since practical usability is central to AI integration into orthodontic workflow. This study has several limitations. First, manual tracing was used as the reference method despite the absence of a true gold standard in cephalometric analysis. Second, because the second examiner reviewed completed tracings to reach consensus rather than performing an independent set of measurements, formal inter‐examiner reliability could not be calculated, limiting our ability to characterize the reproducibility of the reference method. Intra‐examiner reliability, however, was assessed via repeat tracing and was excellent across all seven variables (ICC 0.976–0.996), indicating highly reproducible measurements by the primary examiner despite this limitation. Third, landmark‐level analysis was not performed. Fourth, complete examiner blinding to group allocation was not possible, as cleft‐related anatomical features were visible on the lateral cephalograms, potentially introducing a small degree of observer bias. Fifth, WebCeph did not provide error messages or failure codes for the 17 unprocessed cases, precluding a cause‐based classification of operational failures and limiting identification of modifiable factors to reduce failure rates in future applications. This exclusion (17/75, 22.6%) also reduced the analyzable subgroups to 35 CUCLP and 23 non‐cleft controls, limiting the precision of the between‐group comparisons in Table 7 and warranting caution in interpreting the lack of statistically significant group differences. Finally, the proportional bias observed for ANB and the gonial angle indicates that a single fixed bias and limits of agreement may not fully capture their disagreement pattern across the full measurement range, a consideration when applying these findings to individual patients with markedly high or low values for these parameters. Conclusion WebCeph showed parameter‐dependent agreement with manual cephalometric tracing, moderate to good for skeletal measurements and poor for the gonial angle, IMPA, and nasolabial angle, alongside a considerable rate of processing failure. Although similar patterns of measurement discrepancies and software failures were observed in both complete unilateral cleft lip and palate (CUCLP) and non‐cleft patients, these findings should be interpreted with caution, as a potential influence of cleft‐related anatomy cannot be excluded. Within the scope of the present study, the evaluated version of WebCeph may serve as a supportive tool in the clinical workflow but should not be considered a substitute for manual cephalometric tracing for the seven cephalometric variables assessed. As only a single AI‐based platform was evaluated, these findings should not be generalized to other AI‐assisted cephalometric systems. Author Contributions Aylar Afshari: conceptualization, methodology, data curation, formal analysis, writing – original draft. Zahra Mohandes: conceptualization, methodology, data curation, writing – original draft. Shabnam Ajami: supervision, visualization, project administration, writing – review and editing, resources. Funding The authors have nothing to report. Ethics Statement This study was approved by the Ethics Committee (IR.SUMS.DENTAL. REC.1404.034), conducted in accordance with the Declaration of Helsinki, and was reported according to the STARD guideline. Consent All participants were minors; therefore, written informed consent for publication was obtained from their legal guardians. Conflicts of Interest The authors declare no conflicts of interest. Transparency Statement Aylar Afshari affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained. Supporting information Supporting File