Deep learning–based automatic adenoid segmentation and a novel volume-based index for adenoid hypertrophy assessment

Abstract

Background

Adenoid hypertrophy is a common cause of pediatric obstructive sleep apnea (OSA), which can impair cognitive development and affect craniofacial development. Given that adenoids typically regress with age and may respond to oral appliance therapy, early identification, clear orthodontic indications, and timely referral are essential for effective management. A major challenge in this process is accurately assessing adenoid volume and quantifying nasopharyngeal obstruction in three dimensions. To address this, we developed a deep learning–based method using cone-beam computed tomography (CBCT) to accurately segment adenoid, and establish a novel quantitative index, the Three-Dimensional Adenoidal–Nasopharyngeal Ratio (3D-AN), for assessing nasopharyngeal airway obstruction.

Methods

This retrospective study included three datasets. Dataset 1 comprised 126 CBCT scans (pre- and post-adenoidectomy) from 63 pediatric OSA patients. Dataset 2 included post-adenoidectomy CBCT scans of 26 pediatric OSA patients. Dataset 3 involved 161 pediatric patients with both polysomnography and CBCT data. A SegResNet-based model was trained on Datasets 1 and 2 (1326 training pairs) to predict postoperative airway morphology following adenoidectomy using preoperative scans. By inputting preoperative airway data, the model predicts postoperative morphology, and the volumetric difference represents the adenoid. 3D-AN ratio was then computed from adenoid-to-nasopharyngeal volume. Dataset 3 was used to investigate the relationship between the 3D-AN and pediatric OSA.

Results

The segmentation model achieved a Dice similarity coefficient of 0.88 and a relative volume error of 0.09. Preliminary analysis showed that a 3D-AN ratio below 0.18 was not significantly correlated with pediatric OSA, while values above 0.18 correlated significantly with OSA (r = 0.56, P < 0.01).

Conclusions

This automated method enables accurate adenoid segmentation. The proposed 3D-AN, based on volumetric ratios, holds potential for early detection of adenoid-related airway obstruction and for supporting future research in adenoid assessment and clinical management.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12903-026-07675-2.

Introduction

Pediatric obstructive sleep apnea (OSA) is a breathing disorder caused by upper airway obstruction, which can impact cognitive function, craniofacial development, and daily life in children [1]. Among the many risk factors for pediatric OSA, adenoid hypertrophy is considered one of the most significant [2]. This condition is of particular concern to dental professionals, especially orthodontists, as it is associated with abnormal craniofacial growth [3, 4]. While severe cases of OSA are typically diagnosed early by pediatricians, mild to moderate cases are often discovered during routine dental check-ups or orthodontic evaluations in the mixed dentition stage [5]. Of growing interest is recent research suggesting that oral appliances may help alleviate pediatric OSA symptoms and even reduce the degree of adenoid hypertrophy [6–8]. However, a major challenge faced by dental clinicians and researchers is how to accurately assess adenoid volume and the extent of nasopharyngeal airway obstruction. Although nasopharyngoscopy can accurately assess the degree of nasopharyngeal obstruction, it relies on the clinical judgment of otolaryngologists and is not readily available to dental practitioners [9, 10]. Lateral cephalometric radiographs, while commonly used, are limited by their two-dimensional (2D) nature, often underestimating adenoid volume [11, 12]. Magnetic Resonance Imaging (MRI) provides clear imaging of soft tissues but is costly and not easily accessible. While cone-beam computed tomography (CBCT) is widely used in dentistry for three-dimensional (3D) airway imaging, it is theoretically limited in soft tissue visualization, thus constraining its application in adenoid evaluation [13, 14].

Notably, however, nasopharyngoscopy is used to assess the loss of airway space due to adenoid hypertrophy, while adenoidectomy is performed to restore the partially obstructed nasopharyngeal airway. As a result, clinicians focus primarily on the volume of adenoids to be removed in adenoidectomy. If the nasopharyngeal airway shape after adenoidectomy can be predicted from a pre-treatment CBCT scan, then the difference between the preoperative and predicted postoperative morphology represents the adenoids, providing their 3D information. Advances in artificial intelligence have made this predictive analysis feasible. Therefore, the objectives of this study are as follows:

1. To apply deep learning techniques to analyze 3D morphological changes in the nasopharyngeal airway before and after adenoidectomy, and to develop a predictive model. This model aims to simulate the postoperative shape of the posterior nasopharyngeal wall based on preoperative morphology, thereby enabling automatic segmentation of the adenoids.

2. To establish an index for quantifying the degree of nasopharyngeal airway obstruction, based on airway volume ratios.

3. To explore the relationship between this obstruction index and pediatric OSA.

Methods

Participants

This retrospective study was approved by the Ethics Committee of Shanghai Stomatological Hospital, Fudan University (No. [2024]040), and conducted in accordance with the Declaration of Helsinki. Informed consent was waived by the committee mentioned above, and all data were fully anonymized before the analysis was conducted. All patients underwent imaging as part of their clinical treatment needs. During CBCT acquisition, patients were guided into a natural head position using mechanisms equipped with three-directional laser beams.

This retrospective study reviewed medical records of all patients aged 6–12 treated between 2018 and 2024 at our hospital, including cases from dental prevention centers across four administrative districts in Shanghai. No restrictions were placed on patients’ treatment locations to enhance sample representativeness. Patients with documented adenoid hypertrophy or OSA and relevant medical histories, were included. Based on medical history and data completeness, patients were assigned to the following three datasets.

Dataset 1 and 2 were used to build the prediction model. Dataset 1 included 63 patients with 126 CBCT scans taken both before adenoidectomy and at least six months after surgery, with a maximum interval of one year between scans. There was no recurrence after surgery. No craniofacial surgery or orthodontic treatment occurred between the two scans. CBCTs were required to cover at least the nasopharyngeal and oropharyngeal airways. Patients in Dataset 2 had post-adenoidectomy CBCT scans taken at least 6 months after the surgery, with no recurrence during the postoperative follow-up period. There was no overlap between patients in Datasets 1 and 2.

To investigate the association between the degree of nasopharyngeal airway obstruction and pediatric OSA, we independently reviewed again the whole database comprising children with documented adenoid hypertrophy or OSA. This review resulted in Dataset 3, which included only patients who had not undergone adenoidectomy or orthodontic treatment. The patients had both polysomnography (PSG) records and large-field CBCT scans. The interval between CBCT and PSG was less than one month, and no adenoidectomy, otolaryngologic treatment, craniofacial surgery, or orthodontic treatment occurred during this period. Dataset 3 also had the following exclusion criteria to control for potential confounding factors: severe craniofacial deformities (e.g., cleft lip and palate), a history of craniofacial surgery, chronic pediatric diseases (e.g., asthma), central sleep apnea, genetic syndromes, or pathological obesity. Finally, 161 patients were included. Patients in Database 1 and 2 were entirely excluded due to non-compliance with these criteria. All PSGs were conducted using the same monitoring system (Grael, Condair, Australia); details are provided in the appendix.

Deep learning model for automatic adenoid segmentation

In the data preprocessing stage, as shown in Fig. 1, a threshold-based automatic segmentation method was first used to segment the airway from all CBCT images in Datasets 1 and 2. The segmentation region was confined between the coronal plane at the posterior nasal spine (PNS) and the axial plane at the lower margin of the soft palate. Then, the true annotations for the adenoid were obtained using Dataset 1. Specifically, for one patient’s preoperative and postoperative CBCT images in Dataset 1, initial registration was performed based on anterior cranial base features. A second voxel-similarity-based registration was then conducted, focusing on aligning the posterior nasopharyngeal wall in areas without adenoid hypertrophy. Finally, the region enclosed by the postoperative posterior nasopharyngeal wall and the preoperative posterior nasopharyngeal wall in the adenoid hypertrophy area was defined and labeled as the true adenoid. Following this process, Dataset 1 generated 63 preoperative airway labels with corresponding adenoid labels, while Dataset 2 produced 26 postoperative airway labels.

Fig. 1.

Illustration of data preprocessing and the Three-Dimensional Adenoidal-Nasopharyngeal Ratio (3D-AN). A and B show the data preprocessing, where the red image from post-adenotonsillectomy, overlaid on the gray image from pre-adenotonsillectomy. C illustrates the 3D-AN. The region of interest was bounded anteriorly by the coronal plane passing through the posterior nasal spine (PNS) and inferiorly by the axial plane at the lower margin of the first cervical vertebra. The 3D-AN was defined as the ratio of adenoid volume to the ideal airway volume. The ideal airway volume was calculated by summing the volume of the partially obstructed preoperative airway (highlighted in red segmentation within the box in panel C) and the adenoid volume (highlighted in blue segmentation within the box in panel C)

As shown in Fig. 2, during the deep learning registration model construction and training stage, a model based on the SegResNet framework was developed. The core of the model is to learn the dense deformation field (DDF) of voxel displacement vectors required to align the moving images with the fixed images. The fixed images are the airway segmentation obtained from preoperative CBCT in Dataset 1, while the moving images are the airway segmentation obtained from all CBCT in Dataset 2. To enhance model performance, data augmentation and multi-atlas registration strategies were implemented. Five-fold cross-validation was employed, where the airway segmentations from Dataset 1 and the corresponding adenoid segmentations were randomly divided into five approximately equal groups, with each group containing 12 or 13 samples. In each fold, one group was selected as the test set, while the remaining four groups were used for training. As a result, during training, there were 51 fixed images and 26 moving images, which were randomly paired and input into the model as two channels, creating a total of 1,326 training pairs (51 × 26 = 1,326) to increase the sample size. During training, true adenoid annotations, corresponding to the fixed images, were integrated into the loss function to constrain the learning and prediction process. This constraint offered a dual benefit: it not only assisted the model in locating adenoid hypertrophy regions but also ensured precise alignment outcomes. Specifically, when moving images (airways after adenoidectomy from Dataset 2) were aligned with fixed images (preoperative airway from Dataset 1), non-hypertrophic regions (regions outside the true adenoid mask of preoperative airway from Dataset 1) achieved perfect alignment, while the adenoid hypertrophy regions retained the post-surgical morphology. This process completed the restoration of the partially obstructed airway. The model was trained using the Adam optimizer (initial learning rate: 1 × 10⁻⁴) with an input size of 128 × 128 × 128. The loss function incorporated a similarity constraint based on Dice and a deformation field smoothness regularization term. Detailed model parameters and structural information can be found in the appendix.

Fig. 2.

Deep learning network structure, training, and prediction process. A illustrates the structure of the deep learning network, along with the training and prediction process. The content within the dashed box represents the training stage. After training is completed, the model is deployed for inference, where the ‘input image’ refers to the image whose adenoid needs to be segmented. B shows the comparison between the ground truth and the model’s predictions

The trained model’s subsequent inference process utilized the same multi-atlas registration strategy. In this phase, a single airway image demonstrating untreated adenoid hypertrophy served as the fixed image. All moving images (the entirety of Dataset 2, which had been used in the training stage) were then sequentially registered to this reference image. Each registration generated two outputs: the DDF and the predicted airway label of the moving image after DDF-induced deformation. The predicted airway label represented the post-adenoidectomy airway morphology simulated by the model based on the input hypertrophied airway. By registering each reference airway in Dataset 2 to this input hypertrophied airway, the process yielded a collection of simulated post-adenoidectomy airway morphologies. Finally, all these predicted airway labels were aggregated through voxel-wise averaging to yield the definitive predicted airway morphology. Subsequently, the adenoid morphology and volume were determined by applying a Boolean subtraction of the input pre-operative airway from the predicted post-adenoidectomy airway.

Three-Dimensional Adenoidal-Nasopharyngeal ratio (3D-AN)

Based on the adenoid volume obtained from the above model, we introduced the Three-Dimensional Adenoidal–Nasopharyngeal Ratio (3D-AN), a quantitative index designed to assess the degree of nasopharyngeal airway obstruction in three dimensions. This index was defined as the ratio of the adenoid volume to the total volume of the ideal (unobstructed) airway. The region of interest was bounded anteriorly by the coronal plane passing through the PNS and inferiorly by the axial plane at the lower margin of the first cervical vertebra (Fig. 1C). The ideal airway volume, defined as the expected airway space following adenoidectomy, was calculated by adding the volume of the partially obstructed preoperative airway (red segmentation in Fig. 1C) to the adenoidal volume (blue segmentation in Fig. 1C).

Statistical analysis and evaluation criteria

We evaluated model performance using five-fold cross-validation. The Dice coefficient was used to assess morphological differences between the predicted and the truth, while the relative volume error (RVE) measured differences in volume. Details regarding the sample size for model training and the testing procedures are provided in the appendix.

To compare the model with human experts, we assembled an expert panel consisting of one junior expert (X.Z., over 5 years of experience in pediatric OSA management), one intermediate expert (YY.L., over 10 years of experience), and two senior experts (BJ.Z. and YH.L., over 20 years of experience). The panel manually labeled the adenoids on the preoperative CBCT scans in Dataset 1, and the final labels were obtained by averaging their annotations. Each annotation was reviewed layer by layer and required unanimous approval from the expert panel. The inter-rater agreement for adenoid segmentation, measured by the Dice coefficient, was 0.82 ± 0.05. The Interclass Correlation Coefficient (ICC) for volume was 0.93. The intra-rater agreement, however, showed a Dice of 0.89 ± 0.05 and an ICC for volume higher than 0.96.

The optimal model trained above was applied to Dataset 3 for automated processing to obtain 3D-AN. Apnea-Hypopnea Index (AHI) values were extracted from the corresponding PSG data in Dataset 3. To explore the relationship between 3D-AN and AHI, a scatter plot was first generated. A potential nonlinear association between the two variables was observed. To assess this potential nonlinearity, restricted cubic spline regression with three knots positioned at the 25th, 50th, and 75th percentiles of 3D-AN was performed. The model was adjusted for age, sex, obesity status, and the minimum cross-sectional areas at the velopharyngeal, oropharyngeal, and laryngopharyngeal levels. Piecewise linear regression was then applied to identify possible inflection points. We compared the piecewise model with the standard linear model using the partial F-test, Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). Based on the identified inflection point, Dataset 3 was divided into two subgroups. The Pearson correlation between 3D-AN and AHI was then separately evaluated within each subgroup.

Receiver operating characteristic (ROC) curve analysis was conducted in Dataset 3 to evaluate the screening performance of the 3D-AN index its cut-off value for diagnosing pediatric obstructive sleep apnea (OSA). We assessed the diagnostic ability at two distinct severity thresholds: OSA (AHI ≥ 1 event/h) and moderate-to-severe OSA (AHI ≥ 5 events/h). The area under the curve (AUC), sensitivity, and specificity were calculated. To validate the clinical value of the proposed index, we compared the 3D-AN with established metrics. Due to the retrospective study design, original nasopharyngoscopy records were unavailable. We utilized the four-grade adenoidal scaling system on CBCT images to replace the grading results obtained from nasopharyngoscopy, as previous studies have demonstrated high concordance between each other [15]. Additionally, the traditional 2D Adenoidal-Nasopharyngeal (A/N) ratio was measured on lateral cephalograms [16]. The diagnostic performance (AUC) of the 3D-AN, the four-grade scaling, and the 2D A/N ratio was compared for both OSA severity thresholds (AHI ≥ 1 and AHI ≥ 5). DeLong’s test was utilized to assess the statistical significance of the differences between the AUCs of these methods. Adenoidal grading and the 2D A/N ratio measurement were conducted by two trained examiners. The ICCs were 0.88 and 0.96, respectively, confirming the consistency of the measurements.

The neural network and related automated computations were performed by Python 3.12, based on PyTorch 2.4 and MONAI 1.4. Statistical analyses were performed by R 4.3. All reported p-values were two-tailed, with statistical significance set at p < 0.05.

Results

Dataset 1 included 30 male and 33 female patients. As shown in the Table 1, although there were some morphological differences between model predictions and expert annotations (expert Dice: 0.92 ± 0.03, model Dice: 0.88 ± 0.05), their volumetric agreement was very close (expert RVE: 0.07 ± 0.04, model RVE: 0.09 ± 0.07). In terms of volume, the model’s predictions differed from the ground truth by only about 9%, resulting in minimal impact on the final 3D-AN values (ground truth 3D-AN: 0.23 ± 0.15, model 3D-AN: 0.22 ± 0.15). Figure 2b illustrates the comparison between the model predictions and the ground truth. The predicted results closely approximate the ground truth, demonstrating that the performance has achieved a high level of accuracy.

Table 1.

Comparison of accuracy among ground truth, expert annotation and deep learning model prediction

	Ground Truth	Expert	SegResNet
Dice	-	0.92 ± 0.03	0.88 ± 0.05
RVE	-	0.07 ± 0.04	0.09 ± 0.07
Vad (cm3)	1.58 ± 1.19	1.67 ± 1.29	1.49 ± 1.10
3D-AN	0.23 ± 0.15	0.24 ± 0.16	0.22 ± 0.15

RVE Relative Volume Error, 3D-AN Three-Dimensional Adenoidal-Nasopharyngeal Ratio, Vad Volume of adenoid

The baseline characteristics of Dataset 3 are presented in Table 2. As illustrated in Fig. 3, restricted cubic spline analysis revealed a nonlinear association between AHI and 3D-AN, with an inflection point occurring near the median 3D-AN value of 0.18 ((P for non-linearity < 0.001). The piecewise linear regression model demonstrated a superior fit compared to the standard linear model, evidenced by a higher adjusted R² (0.39 vs. 0.33) and lower information criteria (AIC: 801.24 vs. 813.87; BIC: 832.05 vs. 841.61). The partial F-test further confirmed the statistical significance of the threshold effect (P < 0.001). Below the threshold (3D-AN ≤ 0.18), piecewise linear regression analysis indicated no significant association between 3D-AN and AHI (β = 0.19, 95% CI: -9.12-9.49, P = 0.97). This finding aligns with the lack of correlation observed in this subgroup (r = 0.13, P = 0.26). Conversely, above the threshold (3D-AN > 0.18), a marked increase in the regression slope was observed (β = 26.46, 95% CI: 12.72–40.21, P < 0.001), accompanied by a significant positive correlation (r = 0.56, P < 0.01).

Table 2.

Baseline Characteristics^a

Characteristic	Patients (N = 161)
Age (y)	9.03 ± 1.84
Male sex - no.(%)	86 (53)
BMI	17.50 ± 3.22
Overweight - no.(%)b	32 (20)
AHI	1.93 (2.58)
3D-AN	0.18 (0.16)
VPA (cm2)	1.22 ± 0.67
OPA (cm2)	1.24 ± 0.69
LPA (cm2)	1.31 ± 0.66

a. Normal distribution data are presented as mean ± standard deviation; non-normal distribution data are presented as median (interquartile range). AHI apnea-hypopnea index, BMI body mass index, VPA velopharyngeal minimum cross-sectional area – the airway between the posterior nasal spine plane and the inferior plane of the soft palate, OPA oropharyngeal minimum cross-sectional area – the airway between the inferior plane of the soft palate and the apex of the epiglottis, LPA laryngopharyngeal minimum cross-sectional area – the airway between the plane of the epiglottic apex and the inferior border of the C4 vertebra

b. Overweight was defined as a BMI at or above the 95th percentile

Fig. 3.

Results of restricted cubic splines and piecewise linear regression. A shows the results of the restricted cubic spline. B displays the results of the piecewise linear regression

Regarding the identification of general pediatric OSA (AHI ≥ 1 event/h), the 3D-AN yielded an AUC of 0.56 (Appendix eFigure 3 A). At the determined cut-off value of 0.18, the sensitivity and specificity were 0.52 and 0.58, respectively. Conversely, for the detection of moderate-to-severe OSA (AHI ≥ 5 events/h), the discriminative ability of the 3D-AN improved substantially, achieving an AUC of 0.77 (Appendix eFigure 3B). At the same cut-off of 0.18, the sensitivity rose to 0.80, with a specificity of 0.56. In the comparative analysis, when defining OSA at the threshold of AHI ≥ 1 event/h, the AUCs for the A/N ratio and the four-grade adenoid scale were 0.51 and 0.55, respectively (Appendix eFigure 3 C). DeLong’s test revealed no statistically significant differences between these traditional metrics and the 3D-AN (A/N ratio vs. 3D-AN: P = 0.08, 95% CI -0.01–0.12; Four-grade scale vs. 3D-AN: P = 0.83, 95% CI -0.05–0.07). However, for identifying moderate-to-severe OSA (AHI ≥ 5 events/h), the 3D-AN demonstrated superior performance (Appendix eFigure 3D). The AUCs for the A/N ratio and the four-grade scale were 0.68, both of which were significantly lower than that of the 3D-AN (A/N ratio vs. 3D-AN: P = 0.03, 95% CI 0.01–0.16; Four-grade scale vs. 3D-AN: P = 0.02, 95% CI 0.01–0.17).

Discussion

In this study, a neural network was developed to automatically segment the adenoids. By inputting airway data with adenoid hypertrophy, the model predicted the morphology after adenoidectomy, and the volumetric difference between the two was adenoid volume. Based on the adenoid volume, we proposed the Three-Dimensional Adenoidal-Nasopharyngeal Ratio (3D-AN), defined as the ratio of adenoid volume to the expected nasopharyngeal airway volume. Preliminary analysis indicated that when 3D-AN was below 0.18, no significant association with AHI was observed. However, when the 3D-AN exceeded 0.18, it showed a moderate correlation with AHI.

When otolaryngologists evaluate the nasopharyngeal airway using nasopharyngoscopy, they rely on clinical experience to mentally reconstruct the partially obstructed airway, estimating the degree of obstruction by judging the proportion of the remaining airway in the visual field [17]. This implies that clinicians focus on the volumetric loss caused by adenoid hypertrophy—that is, the volume of adenoid tissue removed during adenoidectomy, reflected by the morphological difference between the preoperative and predicted postoperative airway. Therefore, although CBCT imaging cannot visualize soft tissues as clearly as MRI, if the postoperative posterior nasopharyngeal wall can be accurately predicted, segmentation of the adenoid becomes feasible. On the other hand, since the nasopharynx is adjacent to the cranial base and cervical spine, it is structurally more stable than other soft tissue regions such as the oropharynx [18]. This stability, combined with the consistent head positioning required during oral CBCT acquisition, enhances the feasibility of image registration across different time points [19, 20]. Therefore, we can obtain the ground truth label for the adenoid from Dataset 1.

In this study, all subjects in Dataset 1 had a confirmed clinical diagnosis of adenoid hypertrophy and had undergone adenoidectomy. However, as illustrated in Fig. 1, the segmented adenoid tissue derived from pre- and post-operative CBCT registration may not appear extensively hypertrophic upon visual inspection, with a mean ground truth volume of 1.58 ± 1.19 cm³. This value is notably lower than the MRI-based normative ranges (3.19 to 4.82 cm³) reported in previous literature [21]. This difference stems from the distinct definitions of measured volume: while MRI captures the “total anatomical volume” based on soft-tissue boundaries, our CBCT-based approach quantifies the “effective obstructive volume”. The latter volume represents the net airway space regained following adenoidectomy, reflecting the specific portion of tissue that contributed to airway obstruction. This focus on the obstructive portion aligns with clinical practice in dentistry and otolaryngology, where the primary concern is the degree of functional airway narrowing rather than the absolute size of the adenoid mass.

The prediction of the postoperative posterior nasopharyngeal wall essentially derived from learning the deformation field patterns of the nasopharyngeal airway changes before and after surgery. For such nonlinear patterns, artificial intelligence, particularly deep learning networks, offers significant advantages [22]. Previous studies have applied this technology to tasks such as maxillary clefts completion and tumor radiation changes prediction [23–25]. Considering the scarcity of samples, we adopted several strategies to achieve high performance with small sample training while avoiding overfitting. First, we reduced noise through preprocessing steps like pre-segmentation and cropping. Second, data augmentation was applied to enhance model performance. Finally, both during training and prediction, we utilized a multi-reference image registration strategy, where multiple postoperative CBCT images were used as moving images, and were alternately registered with the fixed image (i.e., the input preoperative airway). The results from these multiple registrations were averaged. During the training phase, the method significantly multiplied the number of samples through the random pairing of fixed and moving images. This strategy resulted in 1,326 training pairs, despite the inherent limitations in the sample sizes of Dataset 1 and Dataset 2. This approach is critically important because, in medical imaging, the availability of training-quality images is often severely restricted by ethical considerations. For instance, Dataset 1 was meticulously curated from a database of thousands of clinical cases. Consequently, a key challenge in the interdisciplinary field of medicine and artificial intelligence is to achieve optimal performance using a limited number of samples. As a result, our model achieved high accuracy, with a Dice of 0.88 and RVE of 0.09, closely matching the expert group’s results, with minimal impact on the final 3D-AN values. This approach greatly reduced, or even eliminated, the need for manual corrections. Furthermore, the development of artificial intelligence and multimodal prediction models has made personalized medicine increasingly feasible. By predicting the morphology of the nasopharyngeal airway after surgery, this approach lays the foundation for individualized outcome prediction of adenoidectomy.

In our preliminary analysis, we found that when the 3D-AN ratio was below 0.18, the AHI did not appear to be associated with nasopharyngeal obstruction. Only when the ratio exceeds 0.18 did nasopharyngeal obstruction begin to show a significant moderate correlation with AHI, aligning the multifactorial etiology of OSA, where adenoids play an important but not exclusive role. These findings are consistent with previous studies and clinical observations suggesting that adenoid hypertrophy alone may not be a strong predictor of pediatric OSA, particularly when it remains within a certain range [26, 27]. Unlike prior studies that relied on ordinal grading or 2D assessments, our method provides a more precise volumetric evaluation of adenoidal obstruction. This may explain the moderate correlation observed between the 3D-AN ratio and AHI beyond the 0.18 threshold. Furthermore, these findings support the interpretation that in cases when adenoid enlargement remains within the normal range, OSA is more likely to result from other anatomical or functional factors rather than nasopharyngeal narrowing [28]. This may also explain why some children experience persistent symptoms after adenoidectomy—other upper airway obstructions remain unresolved, which may be addressed through orthodontic treatment.

Our analysis revealed that when defining OSA with a threshold of AHI ≥ 1, the 3D-AN ratio yielded a modest AUC of 0.56. This finding is consistent with established literature emphasizing the multifactorial etiology of pediatric OSA, where adenoid hypertrophy is merely one of several contributing factors [1]. Consequently, relying solely on a localized anatomical obstruction marker is insufficient for the diagnosis of OSA in the broad pediatric population. However, the clinical utility of the 3D-AN becomes evident in more severe cases. When defining moderate-to-severe OSA with a threshold of AHI ≥ 5 events/h, the AUC improved to 0.77. Notably, at the cut-off of 0.18, the sensitivity reached 0.80. While the specificity (0.56) remained moderate, the high sensitivity suggests that the 3D-AN (with a cut-off of 0.18) is acceptable as a screening indicator, particularly for orthodontists and pediatric dentists. It serves as an effective screening tool to identify children with a nasopharyngeal obstruction who warrant a priority referral to ENT specialists or sleep physicians for polysomnography and comprehensive evaluation. Furthermore, our study demonstrates that the 3D-AN offers a statistically significant improvement in detecting moderate-to-severe OSA compared to traditional methods, including the four-grade scaling and the cephalometric A/N ratio. This suggests that a 3D volumetric quantification provides a more accurate reflection of the nasal obstruction burden caused by adenoid hypertrophy, as associated with the AHI.

In addition to diagnosing nasopharyngeal airway obstruction, the value of the 3D assessment method proposed in this study extends to evaluating the interaction between adenoid growth and nasopharyngeal function. Pereira et al. [29] noted that the adenoids exhibit dynamic developmental characteristics, with their size typically peaking between the ages of 8 and 16, before gradually shrinking. As highlighted in the series of studies by Perry et al. [21, 30, 31], velopharyngeal closure in children and adolescents primarily occurs between the uvula and the adenoids. Therefore, there is a coordinated developmental mechanism between the soft palate and nasopharyngeal volume to maintain effective velopharyngeal closure and resonance function [21, 30, 31]. For example, during the period of adenoid atrophy, the depth dimension decreases first. Simultaneously, the functional length of the soft palate rapidly increases to compensate for the gradual enlargement of the nasopharyngeal cavity. This sustained contact and pressure may be one of the factors inducing the reduction in adenoid depth [30, 32]. However, this dynamic balance will be interrupted by sudden anatomical changes, such as adenoidectomy. If the soft palate’s ability to stretch after surgery cannot manage the abrupt increase in nasopharyngeal depth, this may raise the risk of hypernasality and velopharyngeal insufficiency (VPI) [30, 33].

This risk highlights the importance of precise 3D evaluation of nasopharyngeal airway for children with cleft lip and palate (CLP), especially those presenting with symptoms of OSA. Although adenoid hypertrophy is a major cause of OSA in CLP children, and adenoidectomy can effectively relieve airway blockage, the procedure may increase the risk of VPI due to the inherent craniofacial developmental abnormalities in these patients [31, 34–37]. This constitutes a therapeutic dilemma in clinical practice: prioritizing airway issues may compromise velopharyngeal closure function, and vice versa. Consequently, treatment plans for CLP patients usually require multidisciplinary team collaboration to weigh the potential benefits against the risks [31]. The clinical value of our method shows fully in this complex scenario: unlike conventional 2D imaging, our approach not only provides a 3D quantification of the current adenoid and airway status but also predicts the nasopharyngeal morphology after adenoidectomy. This capability assists clinicians in pre-determining the potential for velopharyngeal closure following adenoidectomy.

In contrast to the risks mentioned above, orthodontic treatment demonstrates some advantages in the management of pediatric OSA. Orthodontic treatments, such as maxillary arch expansion and mandibular advancement, can increase the total upper airway volume by improving the underlying skeletal structures [26, 38]. Furthermore, existing research indicates that orthodontic treatment may induce a reduction in adenoid volume [7, 8, 39]. Given the tendency for adenoid hypertrophy to spontaneously resolve, orthodontic treatment is a possible and effective non-surgical alternative for mild-to-moderate hypertrophy cases. This is particularly important for OSA children with CLP, as orthodontics is a core component of their sequential treatment and represents a slow, continuous process. This process avoids the risk of VPI caused by abrupt changes in the velopharyngeal passage, thereby effectively relieving the conflict between functional restoration and OSA treatment. However, due to a lack of assessment tools for the adenoid and nasopharyngeal airway, dentists found it difficult to quantify these effects. The method proposed in this study fills this gap by not only providing a powerful tool for relevant research but also helping dentists fully contribute to the multidisciplinary sequential treatment of OSA and CLP.

Clinical implications

The clinical value of this study lies in harnessing the potential of routine dental CBCT data. It offers dental clinics an accessible and accurate 3D tool for assessing adenoid and nasopharyngeal airway obstruction. This effectively overcomes the limitations associated with 2D cephalograms noted by Major et al. [40], as well as the limited availability of MRI and nasopharyngoscopy in dental outpatient settings. This enhances the ability of dentists to visually screen for adenoid hypertrophy, identify high-risk patients, and facilitate timely referrals during routine clinical practice. Furthermore, it provides potential quantitative support for the complex clinical decisions mentioned previously. Specially, when managing the complex “airway-speech” balance in OSA children with CLP and monitoring the long-term effectiveness of non-surgical orthodontic treatment, our method assists clinical practice and research. It does this by obtaining adenoid volume measurements and predicting the postoperative nasopharyngeal spatial relationship, thereby maximizing the role of the dentist in the multidisciplinary sequential treatment of pediatric OSA and related conditions.

Limitation

This study has several limitations. Although CT is highly effective for assessing adenoid size, it should not be performed solely for the evaluation of nasopharyngeal airway obstruction. Radiation safety principles, such as the ALARA principle, must be strictly adhered to. CT imaging should only be used when there is a justified clinical indication for other diseases, and in such cases, adenoid evaluation should remain a secondary consideration. Due to radiation exposure, CBCT should never replace nasoendoscopy. However, if pre-existing CT images are available, they can be utilized for assessing nasopharyngeal airway obstruction. It should be noted that children with severe adenoid hypertrophy often do not present to dental clinics for their primary consultation, leading to a certain degree of selection bias. Consequently, our findings are interpreted within the context of dental clinics populations. Additionally, larger and more diverse samples are needed to further validate the proposed 3D-AN cut-off value.

What’s more, other limitation of this study is the sample size, constrained by the strict inclusion criteria for paired pre- and post-operative scans. To minimize the risk of overfitting, we employed data augmentation, particularly elastic deformation, which is effective for learning invariance in small medical datasets [41]. Furthermore, we adopted a combinatorial multi-atlas strategy to generate training pairs, ensuring the model learns generalized anatomical distributions rather than memorizing specific subjects [42]. The model’s stability was confirmed via 5-fold cross-validation [43]. Regarding generalizability, while the current model parameters are optimized for CBCT, the underlying algorithmic framework is modality-agnostic [44]. Future adaptation to other modalities (e.g., MRI or fluoroscopy) is feasible but would require transfer learning or domain adaptation techniques to address intensity shifts [45]. Finally, while the adenoid segment was validated using paired pre- and post-operative CBCT scans (Dataset 1), further multi-center studies with extensive real-world postoperative validation are warranted to enhance its clinical reliability and robustness.

Conclusions

Using deep learning techniques, we developed a model to automatically segment the adenoids. This model simulates the post-adenoidectomy nasopharyngeal airway based on the preoperative airway. The difference between the preoperative structure and the predicted postoperative morphology represents the adenoid. Based on this technology, we proposed a new metric, the Three-Dimensional Adenoidal-Nasopharyngeal Ratio (3D-AN), defined as a volume-based index to quantify the degree of nasopharyngeal airway obstruction. Preliminary analysis suggested that a moderate association between nasopharyngeal obstruction and AHI is primarily observed only when the 3D-AN exceeds 0.18.

Abbreviations

Acronym

Full Form

OSA

Obstructive Sleep Apnea

MRI

Magnetic Resonance Imaging

3D

Three-Dimensional

2D

Two-Dimensional

CBCT

Cone-beam Computed Tomography

PSG

Polysomnography

PNS

Posterior Nasal Spine

DDF

Dense Deformation Field

3D-AN

Three-Dimensional Adenoidal-Nasopharyngeal Ratio

RVE

Relative Volume Error

ICC

Interclass Correlation Coefficient

AHI

Apnea-Hypopnea Index

AIC

Akaike Information Criterion

BIC

Bayesian Information Criterion

ROC

Receiver Operating Characteristic

AUC

Area Under the Curve

A/N

Adenoidal-Nasopharyngeal

VPI

Velopharyngeal Insufficiency

CLP

Cleft Lip and Palate

Authors’ contributions

XZ and YYL. contributed to conception, design, data acquisition and interpretation, drafted and critically revised the manuscript. XZ and YYL contributed equally. XYW, MXL, and PPW contributed to data acquisition. As corresponding authors, BJZ and YHL contributed to conception, design, data acquisition and interpretation, and critically revised the manuscript. All authors read and approved the final manuscript.

Funding

This research was supported by Shanghai Top Priority Research Center Program, 2023ZZ02009 and National Key Clinical Program on Orthodontics. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Data availability

The related code and data sets generated or analyzed during the current study are not publicly available in order to preserve patient confidentiality but are available from the corresponding authors on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Ethics Committee of Shanghai Stomatological Hospital, Fudan University (No. [2024]040), and conducted in accordance with the Declaration of Helsinki. All data were fully anonymized before the analysis was conducted. The requirement for informed consent was waived, as this study is a retrospective analysis of medical records and was deemed unnecessary according to national regulations, specifically Article 39 of the Ethical Review Measures for Biomedical Research Involving Humans in China. The registration number for this study in the Chinese Clinical Trial Registry is ChiCTR2500105732 (Registration Date: 9 July 2025).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.