Establishment of AI-assisted diagnosis of the infraorbital posterior ethmoid cells based on deep learning

Establishment of AI-assisted diagnosis of the infraorbital posterior ethmoid cells based on deep learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Establishment of AI-assisted diagnosis of the infraorbital posterior ethmoid cells based on deep learning Ting Ni, Xusheng Qian, Qiang Zeng, Yingying Ma, Ziran Xie, Yakang Dai, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6612929/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 21 Jul, 2025 Read the published version in BMC Medical Imaging → Version 1 posted 12 You are reading this latest preprint version Abstract Objective : To construct an AI-assisted model for identifying the infraorbital posterior ethmoid cells (IPECs) based on deep learning using sagittal CT images. Methods : Sagittal CT images of 277 samples with and 142 samples without IPECs were retrospectively collected. An experienced radiologist engaged in the relevant aspects picked a sagittal CT image that best showed IPECs. The images were randomly assigned to the training and test sets, with 541 sides in the training set and 97 sides in the test set. The training set was used to perform a five-fold cross-validation, and the results of each fold were used to predict the test set. The model was built using nnUNet, and its performance was evaluated using Dice and standard classification metrics. Results : The model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the additional set. Precision was 0.965 for the training set and 1.000 for the additional set, while sensitivity was 0.981 and 0.967, respectively. A comparison of the diagnostic efficacy between manual outlining by a less-experienced radiologist and AI-assisted outlining showed a significant improvement in detection efficiency (P < 0.05). The AI model aided correctly in identifying and outlining all IPECs, including 12 sides that the radiologist should improve portraying. Conclusion : Artificial intelligence (AI) models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions. The infraorbital posterior ethmoid cells Artificial intelligence Deep learning Efficient detection Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction The anatomy of the paranasal sinuses is delicate, complex, and highly variable. Among the four groups of paranasal sinuses, the ethmoid cells have the greatest degree of variability [ 1 ]. This variability is reflected in the presence or absence of agger nasi cells [ 2 ], Haller cells and sphenoethmoidal air cells (also called Onodi cells) [ 3 ]. The infraorbital posterior ethmoid cell (IPEC) is a developmental variant of the posterior ethmoid cell (PE) that located between the orbital floor and the posterior part of the maxillary sinus (MS) [ 4 ]. Unlike other types of ethmoid cell, the IPEC is poorly recognized and studied. Two types of IPEC can be defined according to its relationship with the MS and drainage channels. Type I IPECs, the ethmomaxillary sinuses (EMSs), enter the MS and drain the superior nasal meatus (SNM) [ 5 ][ 6 ]. Type II IPECs, also known as retromaxillary air cells (RMCs) [ 7 ], retromaxillary pneumatization of posterior ethmoid cells (RM.PEs) [ 8 ], or retroantral ethmoid cells (RAECs) [ 9 ], follow the lamina papyracea (LP) and stop below the orbit where the LP passes in the inferolateral direction to join the orbit floor. The variability of the complex anatomy of the paranasal sinuses may have a significant impact on drainage in sinusitis [ 4 , 7 ]. If the complexity of the pathogenesis of the sinusitis prevents correct and timely treatment, the condition can progress to a chronic form. The IPECs are easily ignored during initial endoscopic sinus surgery (ESS) in sinusitis patients because of their proximity to the LP and hidden location. It is essential to understand the anatomy of the nasal sinuses and their variations prior to ESS to ensure patient safety. Incomplete resection or opening of the IPECs may lead to the recurrence of sinusitis or the need for further surgery [ 7 , 8 , 10 ]. Preoperative diagnostic methods may help localize the IPECs, leading to successful ESS. Today, computed tomography (CT) is used as an excellent noninvasive tool to help assess sinusitis [ 11 , 12 ]. Recognition of IPECs by radiologists may help guide the choice of clinical surgical approach and improve patient outcomes [ 13 ]. In recent years, artificial intelligence (AI) has been widely used in the study of numerous biological systems in radiology, and some progress has been made in its implementation for investigating the paranasal sinuses and sinusitis. Chowdhury et al. [ 14 ] used a convolutional neural network (CNN) to identify the open and closed states of the osteomeatal complex (OMC) on coronal CT images. Humphries et al. [ 15 ] similarly used a CNN to automatically segment the paranasal sinuses on CT images, providing a more objective and quantitative method for assessing inflammation in chronic sinusitis. Massey et al. [ 16 ] scored CT images using the Lund‒Mackay system and the Global Osteitis Scoring Scale and established a deep learning algorithm for conducting quantitative analyses. Zou et al. [ 17 ] utilized AI to diagnose bone remodelling in chronic maxillary sinusitis. However, few studies have investigated the use of AI for assessing the condition of the IPECs, which may be related to the limitations of their anatomical location, their complexity and variability, and the lack of knowledge of their relationship with sinusitis. This study aims to build a model for identifying the IPECs on the basis of deep learning to help radiologists better recognize this anatomical variant and gain a greater understanding of the clinical relationship between the IPECs and sinusitis. Materials and methods This study was approved by the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University. As no procedures in this study were expected to cause additional adverse reactions or impose risks to the subjects, exemption from informed consent was requested. The requirement for informed consent was waived because of the retrospective study design. General information A total of 419 patients who underwent paranasal sinus CT scans in our department from December 2020 to September 2021 were enrolled in the study and classified into the cell group or the normal group. The exclusion criteria included: (1) age less than 18 years; (2) brain and sinus trauma; (3) a history of nasal or sinus surgery; and (4) benign or malignant sinus tumors. The gold standard for the presence of IPECs was identification by 1 radiologist experienced in identifying these structures. A total of 277 patients (167 males and 110 females; aged from 18 to 77 years, average age 39.7 years), with a total of 423 sinuses (54 on the right side, 77 on the left side, and 146 bilateral) were included in the cell group. A total of 142 patients (67 males and 75 females; aged from 18 to 77 years, average age 39.3 years), with a total of 215 sinuses (39 on the right, 30 on the left, and 73 bilateral) were included in the normal group. The patient selection process is shown in Fig. 1 . CT scanning and reconstruction A volumetric scan was performed using a Philips Ingenuity 64-row spiral CT scanner (The Netherlands), with the scanning baseline parallel to the auditory‒orbital line, spanning from the top of the frontal sinus to the inferior border of the maxillary alveolar process. The scanning parameters were as follows: tube voltage, 120 kV; tube current, 280 mA; field of view (FOV), 17×17cm; layer thickness and layer spacing, 0.67 mm; and matrix, 512×512. Acquisition of sagittal CT images of the IPECs The IPECs are characterized by the following traits: (1) extension toward below the orbital floor; and (2) location between the posterior part of the MS and the orbital floor on the basis of coronal, sagittal and axial CT images. The anterior wall of the greater palatine canal was marked as the posterior wall of the MS (Fig. 2 ). The CT images were opened on the picture archiving and communication system (PACS) workstation and subjected to multiplanar reformation (MPR) for orientation calibration of the axial, sagittal and coronal images. In the cell group, the localization line was moved near the level of the greater palatine canal on the side of the IPECs until it could not be moved outwards beyond the lateral wall of the greater palatine canal. The level in the sagittal plane that best demonstrated the anatomy of the IPECs was located and stored. In the normal group, the localization line was moved in the same manner, and the corresponding image that best demonstrated the anatomy of the PE in the sagittal plane was saved. Image annotation The images of the cell group were exported in DICOM format and imported into 3D Slicer software. A radiologist experienced in the study of the IPECs manually delineated the cells as regions of interest (ROIs) (Fig. 3 ). Two radiologists with more than 10 years of experience in nasal imaging reviewed and verified the ROIs. Among a total of 638 sinuses, 541 were used to train the model, and 97 were used to test the model. The 97 sinuses in the test set were then used to test the performance of the less experienced radiologist alone and with assistance from the constructed AI model. Images from the normal group were exported directly in DICOM format without annotation. Construction and validation of the artificial intelligence (AI) models We used the nnUNet architecture, a deep learning framework specialized in medical image segmentation [ 18 ]. nnUNet automatically adapts to the characteristics of the dataset, such as the size, modality, and number of classes of images, providing a robust and efficient solution for medical image analysis. The architecture is based on the well-known U-Net [ 19 ], but incorporates a set of customizations and optimizations for improving segmentation performance across a variety of datasets. The detailed configuration of the network is given in Table 1 . Table 1 The detailed configuration of nnUNet network. Setting Value Original Median Spacing [1.0, 1.0, 1.0] Original Median Shape [1, 784, 784] Image Reader and Writer SimpleITKIO Transpose Order (Forward) [0, 1, 2] Transpose Order (Backward) [0, 1, 2] Batch Size 4 Patch Size [896, 896] Spacing [1.0, 1.0] Normalization Scheme CTNormalization Use Mask for Normalization No Network Type PlainConvUNet Number of Stages 8 Features per Stage [32, 64, 128, 256, 512, 512, 512, 512] Convolution Operation Type Conv2d Kernel Sizes [[ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ], [ 3 , 3 ]] Strides [[ 1 , 1 ], [ 2 , 2 ], [ 2 , 2 ], [ 2 , 2 ], [ 2 , 2 ], [ 2 , 2 ], [ 2 , 2 ], [ 2 , 2 ]] Number of Convolutions per Stage [ 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ] Number of Convolutions per Decoder Stage [ 2 , 2 , 2 , 2 , 2 , 2 , 2 ] Convolution Bias Enabled Normalization Operation InstanceNorm2d Normalization Parameters {"eps": 1e-05, "affine": true} Activation Function LeakyReLU Activation Function Parameters {"inplace": true} The training process involved fivefold cross-validation to ensure a reliable evaluation. First, the dataset was divided into five subsets. Then, the model was trained five times; each time, a different subset was retained for validation after the model had been trained with the other four subsets. For the prediction phase, an ensemble learning approach was applied, in which the outputs of the five models were averaged to generate the final prediction. In the test phase, the predictions of each fold were aggregated to form a complete dataset for the final validation. The test dataset was independent of the training set to provide a reliable evaluation. To convert the segmentation problem into a classification task, we determined whether a sample is classified as positive or negative on the basis of the presence of positive pixels in the segmentation result. Specifically, if the segmentation result contained any positive pixels, the sample was considered positive; otherwise, it was classified as negative. Both the predicted values and the ground-truth labels were processed using the same approach to ensure consistency. The model's performance was evaluated using standard classification metrics, including precision, sensitivity, specificity, F1 score and accuracy. Comparison of manual and AI model-assisted efficacy in detecting IPECs The less experienced radiologist outlined the IPECs alone on the images in the test set and then on the images that the AI model had already labelled. The two sets of segmentations were then evaluated by an experienced radiologist. For a particular sinus, the less experienced radiologist was considered to have correctly identified the sinuses if the degree of overlap of the two segmentations was greater than 90%. Evaluation metrics and statistical analysis The performance of the segmentation model was evaluated using the Dice coefficient, a widely used metric for assessing the similarity between the predicted and ground truth regions in segmentation tasks that offers a quantitative indication of segmentation accuracy. Notably, the Dice coefficient is computed exclusively for positive samples, defined as those containing pixels in the ground truth mask. This approach is adopted because the metric yields an indeterminate 0/0 result for true negative cases (where both ground truth and prediction masks are empty), and it focuses the evaluation specifically on the model's ability to accurately segment existing target regions. It is particularly useful for imbalanced datasets in which the ROIs may be much smaller than the entire image. The formula for the Dice coefficient is given by formula (1), where A and B represent the predicted and true segmentation masks, respectively. The value of the Dice coefficient ranges from 0 (no overlap) to 1 (perfect overlap). $$\:Dice=\frac{2\times\:\mid\:A\cap\:B\mid\:}{\mid\:A\mid\:+\mid\:B\mid\:}\varvec{}$$ 1 The classification performance of the AI model was evaluated using standard classification metrics including precision, sensitivity, specificity, the F1 score, and accuracy. These metrics are commonly used to assess different aspects of model performance, ensuring a balanced evaluation of classification accuracy, error rates, and the trade-off between various types of misclassifications. The formulas for each metric are shown below. $$\:Precision=\frac{TP}{TP+FP}$$ 2 $$\:Sensitivity=\frac{TP}{TP+FN}$$ 3 $$\:F1-score=2\times\:\frac{Precision\times\:Sensitivity}{Precision+Sensitivity}$$ 4 $$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}\varvec{}\varvec{}$$ 5 Statistical analysis was performed in SPSS 24.0 software. The chi-square test was used to compare the results between the manual annotations and those obtained with AI model assistance. The chi-square statistic was calculated with formula (6), where \(\:{O}_{i}\) represents the observed frequency and \(\:{E}_{i}\) ​ represents the expected frequency for each category. For cases with small sample sizes, Fisher’s exact test was used to determine the significance of the differences. The formula for the Fisher exact test, which is used when comparing two categorical variables in a 2x2 contingency table, is given by (7). A p value less than 0.05 was considered statistically significant. $$\:\chi\:2=\sum\:\frac{{({O}_{i}-{E}_{i})}^{2}}{{E}_{i}}$$ 6 $$\:p=\frac{(a+b)!(c+d)!(a+c)!(b+d)!}{(a+b+c+d)!a!b!c!d!}$$ 7 Results AI model construction and validation In the segmentation task, the model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the test set. After postprocessing, in the training set, the model detected IPECs with a sensitivity of 0.981 (355/362), a specificity of 0.927 (166/179), and an accuracy of 0.963 (521/541). In the test set, the model detected IPECs with a sensitivity of 0.967 (59/61), a specificity of 1.000 (36/36), and an accuracy of 0.979 (95/97). Additional performance metrics, including precision and F1 score, are summarized in Table 2 . Examples of segmentation prediction results, the corresponding ground truths, and their comparisons are shown in Fig. 4 . The confusion matrix of the classification results is shown in Fig. 5 . Table 2 To ensure the rigor and impartiality of the experimental evaluation, a five-fold cross-validation protocol was used in Train Dataset. The mean values and standard deviations were meticulously computed and documented. Cohorts Dice Precision Sensitivity Specificity F1-score Accuracy Train 0.900 0.965 0.981 0.927 0.973 0.963 Test 0.891 1.000 0.967 1.000 0.983 0.979 AI model-assisted efficacy in detecting the IPECs The less experienced radiologist identified and outlined the IPECs alone, and with AI assistance, their performance improved for 12 IPECs including 3 cells that the less experienced radiologist failed to identify at all. All the images could be accurately identified and outlined with assistance from the AI model. A comparison of the diagnostic efficacy metrics between the less experienced radiologist alone and with assistance from the AI model is shown in Table 3 . The efficiency in identifying the IPECs was significantly improved with AI assistance (P < 0.05). Table 3 Comparison of results between manual annotation and manual performance with AI model assistance using test set. Result Manual Manual + AI P value True 85 97 0.000 False 12 0 Total 97 97 Discussion Feasibility of the study The IPECs form when the PEs extend below the orbit to the space between the orbital floor and the posterior part of the MS. Due to their anatomical uniqueness, the IPECs tend to be located in a relatively consistent position, at the posterosuperior corner of the MS [ 8 , 20 ]. This facilitates quick identification of this anatomical variant. The collection of sagittal CT images, the orientation in which these cells can best be visualized, is the first step in constructing an AI model for identifying the IPECs. In the present study, images were selected near the level of the greater palatine canal but no farther than near the lateral wall of the greater palatine canal. The two main reasons are as follows. First, the greater palatine canal adjacent to the posterior part of the MS is more easily observed on sagittal images. Second, the EMS is usually surrounded by five walls according to Liu et al. [ 20 ] The posterior wall of the EMS is the inwards, posterior wall of the MS, normally occupying the upper part of the wall. In a few cases, the EMS completely occupies the inward posterior wall of the MS with excessive extension towards the inferior MS. The lateral wall of the EMS is the inwards posterolateral wall of the MS, the depth of which depends on the degree of EMS pneumatization. The lower outlet of the greater palatine canal is the greater palatine foramen, located in the posterior part of the hard palate. Thus, sagittal CT images can best visualize large IPECs while not completely neglecting small IPECs. Results of the study The complex anatomy of the paranasal sinuses has been shown to significantly impact the degree of sinusitis drainage [ 4 , 7 ]. The IPECs have received increasing attention in recent years in numerous studies of the paranasal sinuses, but their results have had little impact on clinical practice. Moreover, an increasing number of studies have utilized AI for systemic radiological diagnosis, demonstrating that AI systems can play important roles in research on the relationship between nasal anatomy and sinusitis [ 14 – 17 ]. Therefore, our research, which is based on the unique anatomical traits depicted on two-dimensional CT images, established an AI model based on deep learning techniques for quick identification of the IPECs. As a type of paranasal sinus, IPECs are complex and heterogeneous. Thus, the successful establishment of an AI classification model on the basis of two-dimensional CT images could demonstrate the feasibility of future three-dimensional model. The successful performance of our AI-assisted diagnostic model highlights the potential of AI and segmentation models in medical image analysis, particularly for identifying subtle anatomical structures such as the IPECs. Deep learning-based segmentation models, such as the nnUNet framework employed in this study, have demonstrated remarkable capabilities in accurately delineating ROIs on medical images. The high Dice coefficient and precision and recall values achieved in our study further demonstrate the effectiveness of the nnUNet model in segmenting the IPECs. Compared with traditional manual segmentation, AI-driven approaches offer advantages in terms of efficiency, consistency, and objectivity. By automating the time-consuming task of manual delineation, AI models can significantly reduce the workloads of radiologists, allowing them to focus on more complex diagnostic and clinical decision-making tasks. Furthermore, the improved detection rate observed with AI assistance, especially in cases missed by the less experienced radiologist, suggests that these models can improve diagnostic accuracy and potentially reduce errors. Clinical significance and directions for future research Excessive pneumatization of the IPECs can squeeze the drainage channel of the MS. Removing the space between the MS and the pneumatized cell during surgery can improve MS drainage [ 1 ]. There are few studies on the association between the IPECs and chronic rhinosinusitis (CRS). Herzallah et al. [ 8 ] reported that type III (L > 6 mm) PE retromaxillary pneumatization has a relatively high prevalence among residual cells on operated sides prepared for revision ESS and concluded that residual undissected RMCs are a common finding in revision ESS. However, some researchers have argued the opposite. Cao et al. [ 1 ], for example, suggested no correlation between the presence of RMCs and ipsilateral MS opacification or between the degree of RMC lateral extension and the Lund–Mackay score of the ipsilateral MS opacification. It is therefore essential to identify the IPECs and explore their importance in depth. Early detection and intervention can effectively interrupt disease progression and further improve patient outcoms. Thus, the establishment of AI predictive models that could guide decision-making in this field seems poised as the dominant direction of current research [ 21 – 23 ]. In addition to the abovementioned studies investigating the use of AI in sinusitis-related aspects such as anatomical status evaluation and segmentation, quantitative analysis based on scoring, and observation of secondary changes [ 14 – 17 ], other studies have investigated the use of AI for predictive analyses. He et al. [ 24 ] first proposed a multitask deep learning framework for constructing a sinus segmentation model based on CT images, and then combined deep learning and clinical factors to predict CRS recurrence. Lai et al. [ 25 ] proposed a method using multiangle sinus CT images combined with AI to predict CRS endotypes, including eosinophilic and non-eosinophilic CRS with nasal polyps. The findings of these studies encourage future research on the association between IPECs and sinusitis. Limitations While our study demonstrates the promise of AI in identifying the IPECs, it is important to acknowledge the limitations inherent in current AI models and segmentation techniques. These include the reliance on high-quality datasets for training, the potential for bias in datasets, affecting model generalizability, and the need for ongoing validation and refinement to ensure robust performance across diverse patient populations and imaging protocols. Sagittal images can be used to rapidly but not necessarily accurately identify the IPECs, an anatomical variant of the paranasal sinuses with a complex morphology. Liu et al. [ 20 ] reported that several anatomical structures, including the SNM, the sphenoid sinus (SS) and the IPECs (EMS and RM.PE), can appear between the MS and the orbital floor alone or simultaneously. Because of the similarity in location and the fact that the two types of IPECs drain into the SNM, identification of the EMS and RP.PE is sometimes difficult. In the process of data collection, we also found that hyperpneumatized SSs, PEs and SNMs can resemble IPECs on sagittal CT images (Fig. 6 ). Correctly identifying the IPECs is therefore highly demanding and requires high levels of expertise on the part of the radiologist. Future research should focus on addressing these limitations and explore the application of 3D segmentation models to capture the complex morphology of the IPECs more comprehensively. Furthermore, the sample size of this study is relatively small, and all the samples came from our unit. Future multicentre, larger-sample studies that integrate AI-assisted tools into clinical workflows are needed to evaluate the real-world impact of these models on surgical planning and patient outcomes. Conclusion The IPECs, which are located between the orbital floor and the posterior part of the MS, are derived from the PEs and drain into the SNM. The presence of IPECs can sometimes cause great difficulties during ESS. Preoperative identification of this anatomic variant is therefore important. AI models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions. Future studies should investigate the construction of three-dimensional models to address the limitations of 2D-image research and better identify mimics at the location of the IPECs. Additionally, larger-sample studies with greater data quality and multicenter experimental validation are needed toincrease the robustness and generalizability of AI models. This will ensure that the models are reliable and applicable across diverse patient populations and clinical settings. Abbreviations IPEC The infraorbital posterior ethmoid cell AI Artificial intelligence MS maxillary sinus PE The posterior ethmoid cell EMS The ethmomaxillary sinus SNM The superior nasal meatus RMC Retromaxillary air cell RM.PE Posterior ethmoid cell RAEC Retroantral ethmoid cell LP The lamina papyracea ESS Endoscopic sinus surgery CT Computed tomography CNN Convolutional neural network OMC The ostiomeatal complex FOV Field of view MPR Multiplanar reorganization ROI Regions of interest CRS Chronic rhinosinusitis SS The sphenoid sinus Declarations Ethics approval and consent to participate This study was approved by the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University; As no procedures in this study were expected to cause additional adverse reactions or impose risks to the subjects, exemption from informed consent was requested. We could confirm that all methods were performed in accordance with the relevant guidelines and regulations. We could confirm that our study adhered to the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee of the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University, because of the retrospective study design. Consent for publication Not Applicable. Competing interests The authors declare no conflicts of interest. Funding The research was funded by Suzhou Science & Technology Projects(SZS2024007)(to Xusheng Qian). Author Contribution Ting Ni and Xusheng Qian wrote the main manuscript text. Ting Ni, Yingying Ma and Ziran Xie collected imaging data. Xusheng Qian, Qiang Zeng and Yakang Dai analyzed data and prepared figure 4,5 and table 1,2. Ting Ni prepared figures 1-3,6 and table 3. Zigang Che supervised the text. All authors reviewed the manuscript. Acknowledgements Not Applicable. Data Availability Data is provided within the manuscript or supplementary information files. The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. References Cao C, Zhou F, Song Z, Tao Z, Xu Y. Computed Tomography Image Analysis and Clinical Correlations of Retromaxillary Cells. Ear Nose Throat J. 2022;101(7):435–42. Peng Z, Wang Y, Fang Y, Wang Y, Chen X, Fan R, Zhang H, Xie Z, Jiang W. Precision Endonasal Endoscopic Surgery of the Frontal Recess Cells and Frontal Sinus Guided by the Natural Sinus Drainage Pathway. Front Surg. 2022;9:862178. Senturk M, Guler I, Azgin I, Sakarya EU, Ovet G, Alatas N, Tolu I, Erdur O. The role of Onodi cells in sphenoiditis: results of multiplanar reconstruction of computed tomography scanning. Braz J Otorhinolaryngol 2017 Jan-Feb;83(1):88–93. JINFENG L, JINSHENG D, XIAOHUI W, et al. The Pneumatization and Adjacent Structure of the Posterior Superior Maxillary Sinus and Its Effect on Nasal Cavity Morphology[J]. Med Sci Monit. 2017;23:4166–74. Ozcan KM, Selcuk A, Oruk V, Sarikaya Y, Dere H. Ethmomaxillary sinus. Eur Arch Otorhinolaryngol. 2008;265(2):185–8. Sirikçi A, Bayazit YA, Bayram M, Kanlikama M. Ethmomaxillary sinus: a particular anatomic variation of the paranasal sinuses. Eur Radiol. 2004;14(2):281–5. KUAN EC, MALLEN-ST CLAIR J, FREDERICK JW, et al. Significance of undissected retromaxillary air cells as a risk factor for revision endoscopic sinus surgery[J]. Am J Rhinol Allergy. 2016;30(6):448–52. HERZALLAH IR, SAATI FA, MARGLANI OA, et al. Retromaxillary Pneumatization of Posterior Ethmoid Air Cells: Novel Description and Surgical Implications[J]. Otolaryngol Head Neck Surg. 2016;155(2):340–6. CHAPURIN N, HONEYBROOK A, JOHNSON S, et al. Radiographic Characterization of the Retroantral Ethmoid Cell[J]. Int Forum Allergy Rhinol. 2016;6(12):1315–8. LEVINE CG, CASIANO RR. Revision Functional Endoscopic Sinus Surgery[J]. Otolaryngol Clin North Am. 2017;50(1):143–64. Che Z, Zhang Q, Zhao P, Lv H, Ding H, Li J, Wang H, Zhang P, Ji H, Zou C, Wang Z. Computed Tomography Evaluation of Unilateral Chronic Maxillary Sinusitis With Osteitis. Ear Nose Throat J. 2023;102(5):NP237–44. Snidvongs K, Sacks R, Harvey RJ. Osteitis in Chronic Rhinosinusitis. Curr Allergy Asthma Rep. 2019;19(5):24. Dahal P, Parajuli S, Maharjan S, Adhikari G, Upadhyaya RP, Ghimire S, Dhakal N. Critical anatomical variants in preoperative computed tomography of paranasal sinuses in a tertiary care center: a cross-sectional study. Ann Med Surg (Lond). 2025;87(4):1909–17. Chowdhury NI, Smith TL, Chandra RK, Turner JH. Automated classification of osteomeatal complex inflammation on computed tomography using convolutional neural networks. Int Forum Allergy Rhinol. 2019;9(1):46–52. Humphries SM, Centeno JP, Notary AM, Gerow J, Cicchetti G, Katial RK, Beswick DM, Ramakrishnan VR, Alam R, Lynch DA. Volumetric assessment of paranasal sinus opacification on computed tomography can be automated using a convolutional neural network. Int Forum Allergy Rhinol. 2020;10(11):1218–25. Massey CJ, Ramos L, Beswick DM, Ramakrishnan VR, Humphries SM. Clinical Validation and Extension of an Automated, Deep Learning-Based Algorithm for Quantitative Sinus CT Analysis. AJNR Am J Neuroradiol. 2022;43(9):1318–24. Zou C, Ji H, Cui J, Qian B, Chen YC, Zhang Q, He S, Sui Y, Bai Y, Zhong Y, Zhang X, Ni T, Che Z. Preliminary study on AI-assisted diagnosis of bone remodeling in chronic maxillary sinusitis. BMC Med Imaging. 2024;24(1):140. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5–9, 2015, proceedings, part III 18. Springer international publishing, 2015: 234–241. Liu J, Dai J, Wen X, Wang Y, Zhang Y, Wang N. Imaging and anatomical features of ethmomaxillary sinus and its differentiation from surrounding air cells. Surg Radiol Anat. 2018;40(2):207–15. Sun R, Zhang M, Yang L, Yang S, Li N, Huang Y, Song H, Wang B, Huang C, Hou F, Wang H. Preoperative CT-based deep learning radiomics model to predict lymph node metastasis and patient prognosis in bladder cancer: a two-center study. Insights Imaging. 2024;15(1):21. Zhang Y, Zhao J, Li Z, Yang M, Ye Z. Preoperative prediction of renal fibrous capsule invasion in clear cell renal cell carcinoma using CT-based radiomics model. Br J Radiol. 2024;97(1161):1557–67. Li X, Zhang J, Leng Y, Liu J, Li L, Wan T, Dong W, Fan B, Gong L. Preoperative prediction of histopathological grading in patients with chondrosarcoma using MRI-based radiomics with semantic features. BMC Med Imaging. 2024;24(1):171. He S, Chen W, Wang X, Xie X, Liu F, Ma X, Li X, Li A, Feng X. Deep learning radiomics-based preoperative prediction of recurrence in chronic rhinosinusitis. iScience. 2023;26(4):106527. Lai S, Kang W, Chen Y, Zou J, Wang S, Zhang X, Zhang X, Lin Y. An End-to-End CRSwNP Prediction with Multichannel ResNet on Computed Tomography. Int J Biomed Imaging. 2024;2024:4960630. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 21 Jul, 2025 Read the published version in BMC Medical Imaging → Version 1 posted Editorial decision: Revision requested 16 Jun, 2025 Reviews received at journal 13 Jun, 2025 Reviews received at journal 09 Jun, 2025 Reviews received at journal 03 Jun, 2025 Reviewers agreed at journal 26 May, 2025 Reviewers agreed at journal 24 May, 2025 Reviewers agreed at journal 14 May, 2025 Reviewers invited by journal 13 May, 2025 Editor assigned by journal 13 May, 2025 Editor invited by journal 12 May, 2025 Submission checks completed at journal 10 May, 2025 First submitted to journal 10 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6612929","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":456340419,"identity":"f94b6682-82ff-4e5b-9a5e-56a3fc0a0c91","order_by":0,"name":"Ting Ni","email":"","orcid":"","institution":"Southeast University","correspondingAuthor":false,"prefix":"","firstName":"Ting","middleName":"","lastName":"Ni","suffix":""},{"id":456340420,"identity":"2c7fe7b4-11df-4961-a4db-3165fe713f17","order_by":1,"name":"Xusheng Qian","email":"","orcid":"","institution":"Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Xusheng","middleName":"","lastName":"Qian","suffix":""},{"id":456340421,"identity":"b6095deb-ebf7-49ad-8a27-b17657858ada","order_by":2,"name":"Qiang Zeng","email":"","orcid":"","institution":"Hohai University","correspondingAuthor":false,"prefix":"","firstName":"Qiang","middleName":"","lastName":"Zeng","suffix":""},{"id":456340422,"identity":"78b953b2-a400-4f7d-88ba-60a9f0a41f94","order_by":3,"name":"Yingying Ma","email":"","orcid":"","institution":"Southeast University","correspondingAuthor":false,"prefix":"","firstName":"Yingying","middleName":"","lastName":"Ma","suffix":""},{"id":456340423,"identity":"27760bde-9e1e-4e1e-b6af-0fada8c20b36","order_by":4,"name":"Ziran Xie","email":"","orcid":"","institution":"Southeast University","correspondingAuthor":false,"prefix":"","firstName":"Ziran","middleName":"","lastName":"Xie","suffix":""},{"id":456340425,"identity":"65ab9a77-ff88-44a0-88c9-6b77c7351a33","order_by":5,"name":"Yakang Dai","email":"","orcid":"","institution":"Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yakang","middleName":"","lastName":"Dai","suffix":""},{"id":456340426,"identity":"a3d74d34-c568-4484-b384-b4f8c23a94ed","order_by":6,"name":"Zigang Che","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAz0lEQVRIiWNgGAWjYDACCSBOADHYGxsffiBNC8/hZmMJorVAGOltAjzE6OCf3Xz4w4NfdnnykQ/bgPrt5HQbCFly51iCQWJfcrHh7cS2BwUMycZmBwhoMZDIMUhI7GFO3Dg7sd1AguFA4jbCWvI/HEjsqU/cOPNgmwQPcVpyGBsSfhxOnC/BSKQWiRtpxgyJDccTN/AkAgPZgAi/8M9Ifvzxx5/qxPntxx8+/FBhJ0dQCxgwtgFdCFZpQIxyMPjDwCDfQLTqUTAKRsEoGGkAALQqR0nyhS6QAAAAAElFTkSuQmCC","orcid":"","institution":"Southeast University","correspondingAuthor":true,"prefix":"","firstName":"Zigang","middleName":"","lastName":"Che","suffix":""}],"badges":[],"createdAt":"2025-05-07 14:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6612929/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6612929/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12880-025-01831-w","type":"published","date":"2025-07-21T15:57:58+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82888952,"identity":"079cab1e-d616-4c83-9a4c-fb23a32e8248","added_by":"auto","created_at":"2025-05-16 12:04:12","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":122847,"visible":true,"origin":"","legend":"\u003cp\u003ePatients selection flowchart\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/33c7c4f56b5c6b3dd6414618.jpeg"},{"id":82891296,"identity":"bd3c1cb1-46e9-44c1-9e29-68f11ae01350","added_by":"auto","created_at":"2025-05-16 12:12:13","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":418874,"visible":true,"origin":"","legend":"\u003cp\u003eBilateral EMS. \u003cstrong\u003ea\u003c/strong\u003e the PEs enter into the MS and form the EMS. The inward posterior wall of the MS is occupied by the EMS. The arrows show the greater palatine canal. \u003cstrong\u003eb,c\u003c/strong\u003eRight and left sagittal images of the EMS at the plane of near the greater palatine canal (asterisk).\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/d3f02661ddb666b3e00e1fd6.jpeg"},{"id":82888957,"identity":"37289c94-2098-4310-9a5c-c720807b59a7","added_by":"auto","created_at":"2025-05-16 12:04:12","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":43556,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROI area of IPECs in cell group (green) delineated with 3D Slicer software.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/0275cfe9b7374d60879943e7.jpeg"},{"id":82888968,"identity":"0967b96f-a51f-41b4-83a1-9d6b38978960","added_by":"auto","created_at":"2025-05-16 12:04:13","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":599140,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of partial AI segmentation results\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/563925f4430fd82417c70080.png"},{"id":82888962,"identity":"07e71833-288a-4025-80dc-06b16037999e","added_by":"auto","created_at":"2025-05-16 12:04:13","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":42905,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix Based on Five-Fold Cross-Validation Training-Testing Sets and Independent Test Set\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/62c70dfeae3cdf956f6e233d.png"},{"id":82891294,"identity":"77c8e877-4048-4f4f-8345-c6985d969e56","added_by":"auto","created_at":"2025-05-16 12:12:12","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":375191,"visible":true,"origin":"","legend":"\u003cp\u003eSimilar anatomies appearing between the orbital floor and the MS. a The coronal image of the SNM (arrow). b,c The right sphenoid sinus gasifies into the front of the optic canal. The sagittal image of the SS mimics the IPEC in the position between the orbit and the MS. Asterisk represents the greater palatine canal.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/8df2f24d1689bcee8875f706.jpeg"},{"id":87756857,"identity":"4ddab47a-5058-486c-b548-820fe246a848","added_by":"auto","created_at":"2025-07-28 16:09:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2260912,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6612929/v1/c67f5eb9-fd8a-4787-98a7-d5305e6cd922.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Establishment of AI-assisted diagnosis of the infraorbital posterior ethmoid cells based on deep learning","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe anatomy of the paranasal sinuses is delicate, complex, and highly variable. Among the four groups of paranasal sinuses, the ethmoid cells have the greatest degree of variability [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. This variability is reflected in the presence or absence of agger nasi cells [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], Haller cells and sphenoethmoidal air cells (also called Onodi cells) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The infraorbital posterior ethmoid cell (IPEC) is a developmental variant of the posterior ethmoid cell (PE) that located between the orbital floor and the posterior part of the maxillary sinus (MS) [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Unlike other types of ethmoid cell, the IPEC is poorly recognized and studied. Two types of IPEC can be defined according to its relationship with the MS and drainage channels. Type I IPECs, the ethmomaxillary sinuses (EMSs), enter the MS and drain the superior nasal meatus (SNM) [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e][\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Type II IPECs, also known as retromaxillary air cells (RMCs) [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], retromaxillary pneumatization of posterior ethmoid cells (RM.PEs) [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], or retroantral ethmoid cells (RAECs) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], follow the lamina papyracea (LP) and stop below the orbit where the LP passes in the inferolateral direction to join the orbit floor.\u003c/p\u003e \u003cp\u003eThe variability of the complex anatomy of the paranasal sinuses may have a significant impact on drainage in sinusitis [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. If the complexity of the pathogenesis of the sinusitis prevents correct and timely treatment, the condition can progress to a chronic form. The IPECs are easily ignored during initial endoscopic sinus surgery (ESS) in sinusitis patients because of their proximity to the LP and hidden location. It is essential to understand the anatomy of the nasal sinuses and their variations prior to ESS to ensure patient safety. Incomplete resection or opening of the IPECs may lead to the recurrence of sinusitis or the need for further surgery [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Preoperative diagnostic methods may help localize the IPECs, leading to successful ESS. Today, computed tomography (CT) is used as an excellent noninvasive tool to help assess sinusitis [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Recognition of IPECs by radiologists may help guide the choice of clinical surgical approach and improve patient outcomes [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn recent years, artificial intelligence (AI) has been widely used in the study of numerous biological systems in radiology, and some progress has been made in its implementation for investigating the paranasal sinuses and sinusitis. Chowdhury et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] used a convolutional neural network (CNN) to identify the open and closed states of the osteomeatal complex (OMC) on coronal CT images. Humphries et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] similarly used a CNN to automatically segment the paranasal sinuses on CT images, providing a more objective and quantitative method for assessing inflammation in chronic sinusitis. Massey et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] scored CT images using the Lund‒Mackay system and the Global Osteitis Scoring Scale and established a deep learning algorithm for conducting quantitative analyses. Zou et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] utilized AI to diagnose bone remodelling in chronic maxillary sinusitis. However, few studies have investigated the use of AI for assessing the condition of the IPECs, which may be related to the limitations of their anatomical location, their complexity and variability, and the lack of knowledge of their relationship with sinusitis. This study aims to build a model for identifying the IPECs on the basis of deep learning to help radiologists better recognize this anatomical variant and gain a greater understanding of the clinical relationship between the IPECs and sinusitis.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e This study was approved by the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University. As no procedures in this study were expected to cause additional adverse reactions or impose risks to the subjects, exemption from informed consent was requested. The requirement for informed consent was waived because of the retrospective study design.\u003c/p\u003e \u003cp\u003eGeneral information\u003c/p\u003e \u003cp\u003eA total of 419 patients who underwent paranasal sinus CT scans in our department from December 2020 to September 2021 were enrolled in the study and classified into the cell group or the normal group. The exclusion criteria included: (1) age less than 18 years; (2) brain and sinus trauma; (3) a history of nasal or sinus surgery; and (4) benign or malignant sinus tumors. The gold standard for the presence of IPECs was identification by 1 radiologist experienced in identifying these structures. A total of 277 patients (167 males and 110 females; aged from 18 to 77 years, average age 39.7 years), with a total of 423 sinuses (54 on the right side, 77 on the left side, and 146 bilateral) were included in the cell group. A total of 142 patients (67 males and 75 females; aged from 18 to 77 years, average age 39.3 years), with a total of 215 sinuses (39 on the right, 30 on the left, and 73 bilateral) were included in the normal group. The patient selection process is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eCT scanning and reconstruction\u003c/p\u003e \u003cp\u003eA volumetric scan was performed using a Philips Ingenuity 64-row spiral CT scanner (The Netherlands), with the scanning baseline parallel to the auditory‒orbital line, spanning from the top of the frontal sinus to the inferior border of the maxillary alveolar process. The scanning parameters were as follows: tube voltage, 120 kV; tube current, 280 mA; field of view (FOV), 17\u0026times;17cm; layer thickness and layer spacing, 0.67 mm; and matrix, 512\u0026times;512.\u003c/p\u003e \u003cp\u003eAcquisition of sagittal CT images of the IPECs\u003c/p\u003e \u003cp\u003eThe IPECs are characterized by the following traits: (1) extension toward below the orbital floor; and (2) location between the posterior part of the MS and the orbital floor on the basis of coronal, sagittal and axial CT images. The anterior wall of the greater palatine canal was marked as the posterior wall of the MS (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe CT images were opened on the picture archiving and communication system (PACS) workstation and subjected to multiplanar reformation (MPR) for orientation calibration of the axial, sagittal and coronal images. In the cell group, the localization line was moved near the level of the greater palatine canal on the side of the IPECs until it could not be moved outwards beyond the lateral wall of the greater palatine canal. The level in the sagittal plane that best demonstrated the anatomy of the IPECs was located and stored. In the normal group, the localization line was moved in the same manner, and the corresponding image that best demonstrated the anatomy of the PE in the sagittal plane was saved.\u003c/p\u003e \u003cp\u003eImage annotation\u003c/p\u003e \u003cp\u003eThe images of the cell group were exported in DICOM format and imported into 3D Slicer software. A radiologist experienced in the study of the IPECs manually delineated the cells as regions of interest (ROIs) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Two radiologists with more than 10 years of experience in nasal imaging reviewed and verified the ROIs. Among a total of 638 sinuses, 541 were used to train the model, and 97 were used to test the model. The 97 sinuses in the test set were then used to test the performance of the less experienced radiologist alone and with assistance from the constructed AI model. Images from the normal group were exported directly in DICOM format without annotation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eConstruction and validation of the artificial intelligence (AI) models\u003c/p\u003e \u003cp\u003eWe used the nnUNet architecture, a deep learning framework specialized in medical image segmentation [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. nnUNet automatically adapts to the characteristics of the dataset, such as the size, modality, and number of classes of images, providing a robust and efficient solution for medical image analysis. The architecture is based on the well-known U-Net [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], but incorporates a set of customizations and optimizations for improving segmentation performance across a variety of datasets. The detailed configuration of the network is given in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe detailed configuration of nnUNet network.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSetting\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOriginal Median Spacing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[1.0, 1.0, 1.0]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOriginal Median Shape\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[1, 784, 784]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImage Reader and Writer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSimpleITKIO\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTranspose Order (Forward)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[0, 1, 2]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTranspose Order (Backward)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[0, 1, 2]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBatch Size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatch Size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[896, 896]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpacing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[1.0, 1.0]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormalization Scheme\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCTNormalization\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUse Mask for Normalization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNetwork Type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePlainConvUNet\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Stages\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeatures per Stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[32, 64, 128, 256, 512, 512, 512, 512]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConvolution Operation Type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConv2d\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKernel Sizes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStrides\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Convolutions per Stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Convolutions per Decoder Stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConvolution Bias\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnabled\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormalization Operation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInstanceNorm2d\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormalization Parameters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e{\"eps\": 1e-05, \"affine\": true}\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActivation Function\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLeakyReLU\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActivation Function Parameters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e{\"inplace\": true}\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe training process involved fivefold cross-validation to ensure a reliable evaluation. First, the dataset was divided into five subsets. Then, the model was trained five times; each time, a different subset was retained for validation after the model had been trained with the other four subsets. For the prediction phase, an ensemble learning approach was applied, in which the outputs of the five models were averaged to generate the final prediction. In the test phase, the predictions of each fold were aggregated to form a complete dataset for the final validation. The test dataset was independent of the training set to provide a reliable evaluation.\u003c/p\u003e \u003cp\u003eTo convert the segmentation problem into a classification task, we determined whether a sample is classified as positive or negative on the basis of the presence of positive pixels in the segmentation result. Specifically, if the segmentation result contained any positive pixels, the sample was considered positive; otherwise, it was classified as negative. Both the predicted values and the ground-truth labels were processed using the same approach to ensure consistency. The model's performance was evaluated using standard classification metrics, including precision, sensitivity, specificity, F1 score and accuracy.\u003c/p\u003e \u003cp\u003eComparison of manual and AI model-assisted efficacy in detecting IPECs\u003c/p\u003e \u003cp\u003eThe less experienced radiologist outlined the IPECs alone on the images in the test set and then on the images that the AI model had already labelled. The two sets of segmentations were then evaluated by an experienced radiologist. For a particular sinus, the less experienced radiologist was considered to have correctly identified the sinuses if the degree of overlap of the two segmentations was greater than 90%.\u003c/p\u003e \u003cp\u003eEvaluation metrics and statistical analysis\u003c/p\u003e \u003cp\u003eThe performance of the segmentation model was evaluated using the Dice coefficient, a widely used metric for assessing the similarity between the predicted and ground truth regions in segmentation tasks that offers a quantitative indication of segmentation accuracy. Notably, the Dice coefficient is computed exclusively for positive samples, defined as those containing pixels in the ground truth mask. This approach is adopted because the metric yields an indeterminate 0/0 result for true negative cases (where both ground truth and prediction masks are empty), and it focuses the evaluation specifically on the model's ability to accurately segment existing target regions. It is particularly useful for imbalanced datasets in which the ROIs may be much smaller than the entire image. The formula for the Dice coefficient is given by formula (1), where A and B represent the predicted and true segmentation masks, respectively. The value of the Dice coefficient ranges from 0 (no overlap) to 1 (perfect overlap).\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:Dice=\\frac{2\\times\\:\\mid\\:A\\cap\\:B\\mid\\:}{\\mid\\:A\\mid\\:+\\mid\\:B\\mid\\:}\\varvec{}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe classification performance of the AI model was evaluated using standard classification metrics including precision, sensitivity, specificity, the F1 score, and accuracy. These metrics are commonly used to assess different aspects of model performance, ensuring a balanced evaluation of classification accuracy, error rates, and the trade-off between various types of misclassifications. The formulas for each metric are shown below.\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:Precision=\\frac{TP}{TP+FP}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:Sensitivity=\\frac{TP}{TP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:F1-score=2\\times\\:\\frac{Precision\\times\\:Sensitivity}{Precision+Sensitivity}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\frac{TP+TN}{TP+TN+FP+FN}\\varvec{}\\varvec{}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eStatistical analysis was performed in SPSS 24.0 software. The chi-square test was used to compare the results between the manual annotations and those obtained with AI model assistance. The chi-square statistic was calculated with formula (6), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{O}_{i}\\)\u003c/span\u003e\u003c/span\u003e represents the observed frequency and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{E}_{i}\\)\u003c/span\u003e\u003c/span\u003e​ represents the expected frequency for each category. For cases with small sample sizes, Fisher\u0026rsquo;s exact test was used to determine the significance of the differences. The formula for the Fisher exact test, which is used when comparing two categorical variables in a 2x2 contingency table, is given by (7). A p value less than 0.05 was considered statistically significant.\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:\\chi\\:2=\\sum\\:\\frac{{({O}_{i}-{E}_{i})}^{2}}{{E}_{i}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ7\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\:p=\\frac{(a+b)!(c+d)!(a+c)!(b+d)!}{(a+b+c+d)!a!b!c!d!}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAI model construction and validation\u003c/p\u003e \u003cp\u003eIn the segmentation task, the model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the test set. After postprocessing, in the training set, the model detected IPECs with a sensitivity of 0.981 (355/362), a specificity of 0.927 (166/179), and an accuracy of 0.963 (521/541). In the test set, the model detected IPECs with a sensitivity of 0.967 (59/61), a specificity of 1.000 (36/36), and an accuracy of 0.979 (95/97). Additional performance metrics, including precision and F1 score, are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Examples of segmentation prediction results, the corresponding ground truths, and their comparisons are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The confusion matrix of the classification results is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTo ensure the rigor and impartiality of the experimental evaluation, a five-fold cross-validation protocol was used in Train Dataset. The mean values and standard deviations were meticulously computed and documented.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCohorts\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDice\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.965\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.981\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.927\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.963\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.891\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.983\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAI model-assisted efficacy in detecting the IPECs\u003c/p\u003e \u003cp\u003eThe less experienced radiologist identified and outlined the IPECs alone, and with AI assistance, their performance improved for 12 IPECs including 3 cells that the less experienced radiologist failed to identify at all. All the images could be accurately identified and outlined with assistance from the AI model. A comparison of the diagnostic efficacy metrics between the less experienced radiologist alone and with assistance from the AI model is shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The efficiency in identifying the IPECs was significantly improved with AI assistance (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of results between manual annotation and manual performance with AI model assistance using test set.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResult\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eManual\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eManual\u0026thinsp;+\u0026thinsp;AI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTrue\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eFeasibility of the study\u003c/h2\u003e \u003cp\u003eThe IPECs form when the PEs extend below the orbit to the space between the orbital floor and the posterior part of the MS. Due to their anatomical uniqueness, the IPECs tend to be located in a relatively consistent position, at the posterosuperior corner of the MS [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This facilitates quick identification of this anatomical variant. The collection of sagittal CT images, the orientation in which these cells can best be visualized, is the first step in constructing an AI model for identifying the IPECs. In the present study, images were selected near the level of the greater palatine canal but no farther than near the lateral wall of the greater palatine canal. The two main reasons are as follows. First, the greater palatine canal adjacent to the posterior part of the MS is more easily observed on sagittal images. Second, the EMS is usually surrounded by five walls according to Liu et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] The posterior wall of the EMS is the inwards, posterior wall of the MS, normally occupying the upper part of the wall. In a few cases, the EMS completely occupies the inward posterior wall of the MS with excessive extension towards the inferior MS. The lateral wall of the EMS is the inwards posterolateral wall of the MS, the depth of which depends on the degree of EMS pneumatization. The lower outlet of the greater palatine canal is the greater palatine foramen, located in the posterior part of the hard palate. Thus, sagittal CT images can best visualize large IPECs while not completely neglecting small IPECs.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eResults of the study\u003c/h3\u003e\n\u003cp\u003eThe complex anatomy of the paranasal sinuses has been shown to significantly impact the degree of sinusitis drainage [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. The IPECs have received increasing attention in recent years in numerous studies of the paranasal sinuses, but their results have had little impact on clinical practice. Moreover, an increasing number of studies have utilized AI for systemic radiological diagnosis, demonstrating that AI systems can play important roles in research on the relationship between nasal anatomy and sinusitis [\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Therefore, our research, which is based on the unique anatomical traits depicted on two-dimensional CT images, established an AI model based on deep learning techniques for quick identification of the IPECs. As a type of paranasal sinus, IPECs are complex and heterogeneous. Thus, the successful establishment of an AI classification model on the basis of two-dimensional CT images could demonstrate the feasibility of future three-dimensional model.\u003c/p\u003e \u003cp\u003eThe successful performance of our AI-assisted diagnostic model highlights the potential of AI and segmentation models in medical image analysis, particularly for identifying subtle anatomical structures such as the IPECs. Deep learning-based segmentation models, such as the nnUNet framework employed in this study, have demonstrated remarkable capabilities in accurately delineating ROIs on medical images. The high Dice coefficient and precision and recall values achieved in our study further demonstrate the effectiveness of the nnUNet model in segmenting the IPECs. Compared with traditional manual segmentation, AI-driven approaches offer advantages in terms of efficiency, consistency, and objectivity. By automating the time-consuming task of manual delineation, AI models can significantly reduce the workloads of radiologists, allowing them to focus on more complex diagnostic and clinical decision-making tasks. Furthermore, the improved detection rate observed with AI assistance, especially in cases missed by the less experienced radiologist, suggests that these models can improve diagnostic accuracy and potentially reduce errors.\u003c/p\u003e\n\u003ch3\u003eClinical significance and directions for future research\u003c/h3\u003e\n\u003cp\u003eExcessive pneumatization of the IPECs can squeeze the drainage channel of the MS. Removing the space between the MS and the pneumatized cell during surgery can improve MS drainage [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. There are few studies on the association between the IPECs and chronic rhinosinusitis (CRS). Herzallah et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] reported that type III (L\u0026thinsp;\u0026gt;\u0026thinsp;6 mm) PE retromaxillary pneumatization has a relatively high prevalence among residual cells on operated sides prepared for revision ESS and concluded that residual undissected RMCs are a common finding in revision ESS. However, some researchers have argued the opposite. Cao et al. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], for example, suggested no correlation between the presence of RMCs and ipsilateral MS opacification or between the degree of RMC lateral extension and the Lund\u0026ndash;Mackay score of the ipsilateral MS opacification. It is therefore essential to identify the IPECs and explore their importance in depth.\u003c/p\u003e \u003cp\u003eEarly detection and intervention can effectively interrupt disease progression and further improve patient outcoms. Thus, the establishment of AI predictive models that could guide decision-making in this field seems poised as the dominant direction of current research [\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. In addition to the abovementioned studies investigating the use of AI in sinusitis-related aspects such as anatomical status evaluation and segmentation, quantitative analysis based on scoring, and observation of secondary changes [\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], other studies have investigated the use of AI for predictive analyses. He et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] first proposed a multitask deep learning framework for constructing a sinus segmentation model based on CT images, and then combined deep learning and clinical factors to predict CRS recurrence. Lai et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] proposed a method using multiangle sinus CT images combined with AI to predict CRS endotypes, including eosinophilic and non-eosinophilic CRS with nasal polyps. The findings of these studies encourage future research on the association between IPECs and sinusitis.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eWhile our study demonstrates the promise of AI in identifying the IPECs, it is important to acknowledge the limitations inherent in current AI models and segmentation techniques. These include the reliance on high-quality datasets for training, the potential for bias in datasets, affecting model generalizability, and the need for ongoing validation and refinement to ensure robust performance across diverse patient populations and imaging protocols. Sagittal images can be used to rapidly but not necessarily accurately identify the IPECs, an anatomical variant of the paranasal sinuses with a complex morphology. Liu et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] reported that several anatomical structures, including the SNM, the sphenoid sinus (SS) and the IPECs (EMS and RM.PE), can appear between the MS and the orbital floor alone or simultaneously. Because of the similarity in location and the fact that the two types of IPECs drain into the SNM, identification of the EMS and RP.PE is sometimes difficult. In the process of data collection, we also found that hyperpneumatized SSs, PEs and SNMs can resemble IPECs on sagittal CT images (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). Correctly identifying the IPECs is therefore highly demanding and requires high levels of expertise on the part of the radiologist. Future research should focus on addressing these limitations and explore the application of 3D segmentation models to capture the complex morphology of the IPECs more comprehensively. Furthermore, the sample size of this study is relatively small, and all the samples came from our unit. Future multicentre, larger-sample studies that integrate AI-assisted tools into clinical workflows are needed to evaluate the real-world impact of these models on surgical planning and patient outcomes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe IPECs, which are located between the orbital floor and the posterior part of the MS, are derived from the PEs and drain into the SNM. The presence of IPECs can sometimes cause great difficulties during ESS. Preoperative identification of this anatomic variant is therefore important. AI models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions. Future studies should investigate the construction of three-dimensional models to address the limitations of 2D-image research and better identify mimics at the location of the IPECs. Additionally, larger-sample studies with greater data quality and multicenter experimental validation are needed toincrease the robustness and generalizability of AI models. This will ensure that the models are reliable and applicable across diverse patient populations and clinical settings.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eIPEC The infraorbital posterior ethmoid cell\u003c/p\u003e\u003cp\u003eAI Artificial intelligence\u003c/p\u003e\u003cp\u003eMS maxillary sinus\u003c/p\u003e\u003cp\u003ePE The posterior ethmoid cell\u003c/p\u003e\u003cp\u003eEMS The ethmomaxillary sinus\u003c/p\u003e\u003cp\u003eSNM The superior nasal meatus\u003c/p\u003e\u003cp\u003eRMC Retromaxillary air cell\u003c/p\u003e\u003cp\u003eRM.PE Posterior ethmoid cell\u003c/p\u003e\u003cp\u003eRAEC Retroantral ethmoid cell\u003c/p\u003e\u003cp\u003eLP The lamina papyracea\u003c/p\u003e\u003cp\u003eESS Endoscopic sinus surgery\u003c/p\u003e\u003cp\u003eCT Computed tomography\u003c/p\u003e\u003cp\u003eCNN Convolutional neural network\u003c/p\u003e\u003cp\u003eOMC The ostiomeatal complex\u003c/p\u003e\u003cp\u003eFOV Field of view\u003c/p\u003e\u003cp\u003eMPR Multiplanar reorganization\u003c/p\u003e\u003cp\u003eROI Regions of interest\u003c/p\u003e\u003cp\u003eCRS Chronic rhinosinusitis\u003c/p\u003e\u003cp\u003eSS The sphenoid sinus\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003c/p\u003e\u003cp\u003e This study was approved by the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University; As no procedures in this study were expected to cause additional adverse reactions or impose risks to the subjects, exemption from informed consent was requested. We could confirm that all methods were performed in accordance with the relevant guidelines and regulations. We could confirm that our study adhered to the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee of the ethics committees of Nanjing Tongren Hospital, School of Medicine, Southeast University, because of the retrospective study design.\u003c/p\u003e \u003cp\u003e\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003c/p\u003e\u003cp\u003eNot Applicable.\u003c/p\u003e \u003cp\u003e\u003c/p\u003e\u003cp\u003e \u003c/p\u003e\u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare no conflicts of interest.\u003c/p\u003e \u003cp\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe research was funded by Suzhou Science \u0026amp; Technology Projects(SZS2024007)(to Xusheng Qian).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eTing Ni and Xusheng Qian wrote the main manuscript text. Ting Ni, Yingying Ma and Ziran Xie collected imaging data. Xusheng Qian, Qiang Zeng and Yakang Dai analyzed data and prepared figure 4,5 and table 1,2. Ting Ni prepared figures 1-3,6 and table 3. Zigang Che supervised the text. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNot Applicable.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData is provided within the manuscript or supplementary information files. The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCao C, Zhou F, Song Z, Tao Z, Xu Y. Computed Tomography Image Analysis and Clinical Correlations of Retromaxillary Cells. Ear Nose Throat J. 2022;101(7):435\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng Z, Wang Y, Fang Y, Wang Y, Chen X, Fan R, Zhang H, Xie Z, Jiang W. Precision Endonasal Endoscopic Surgery of the Frontal Recess Cells and Frontal Sinus Guided by the Natural Sinus Drainage Pathway. Front Surg. 2022;9:862178.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSenturk M, Guler I, Azgin I, Sakarya EU, Ovet G, Alatas N, Tolu I, Erdur O. The role of Onodi cells in sphenoiditis: results of multiplanar reconstruction of computed tomography scanning. Braz J Otorhinolaryngol 2017 Jan-Feb;83(1):88\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJINFENG L, JINSHENG D, XIAOHUI W, et al. The Pneumatization and Adjacent Structure of the Posterior Superior Maxillary Sinus and Its Effect on Nasal Cavity Morphology[J]. Med Sci Monit. 2017;23:4166\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOzcan KM, Selcuk A, Oruk V, Sarikaya Y, Dere H. Ethmomaxillary sinus. Eur Arch Otorhinolaryngol. 2008;265(2):185\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSirik\u0026ccedil;i A, Bayazit YA, Bayram M, Kanlikama M. Ethmomaxillary sinus: a particular anatomic variation of the paranasal sinuses. Eur Radiol. 2004;14(2):281\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKUAN EC, MALLEN-ST CLAIR J, FREDERICK JW, et al. Significance of undissected retromaxillary air cells as a risk factor for revision endoscopic sinus surgery[J]. Am J Rhinol Allergy. 2016;30(6):448\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHERZALLAH IR, SAATI FA, MARGLANI OA, et al. Retromaxillary Pneumatization of Posterior Ethmoid Air Cells: Novel Description and Surgical Implications[J]. Otolaryngol Head Neck Surg. 2016;155(2):340\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCHAPURIN N, HONEYBROOK A, JOHNSON S, et al. Radiographic Characterization of the Retroantral Ethmoid Cell[J]. Int Forum Allergy Rhinol. 2016;6(12):1315\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLEVINE CG, CASIANO RR. Revision Functional Endoscopic Sinus Surgery[J]. Otolaryngol Clin North Am. 2017;50(1):143\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChe Z, Zhang Q, Zhao P, Lv H, Ding H, Li J, Wang H, Zhang P, Ji H, Zou C, Wang Z. Computed Tomography Evaluation of Unilateral Chronic Maxillary Sinusitis With Osteitis. Ear Nose Throat J. 2023;102(5):NP237\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSnidvongs K, Sacks R, Harvey RJ. Osteitis in Chronic Rhinosinusitis. Curr Allergy Asthma Rep. 2019;19(5):24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDahal P, Parajuli S, Maharjan S, Adhikari G, Upadhyaya RP, Ghimire S, Dhakal N. Critical anatomical variants in preoperative computed tomography of paranasal sinuses in a tertiary care center: a cross-sectional study. Ann Med Surg (Lond). 2025;87(4):1909\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChowdhury NI, Smith TL, Chandra RK, Turner JH. Automated classification of osteomeatal complex inflammation on computed tomography using convolutional neural networks. Int Forum Allergy Rhinol. 2019;9(1):46\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHumphries SM, Centeno JP, Notary AM, Gerow J, Cicchetti G, Katial RK, Beswick DM, Ramakrishnan VR, Alam R, Lynch DA. Volumetric assessment of paranasal sinus opacification on computed tomography can be automated using a convolutional neural network. Int Forum Allergy Rhinol. 2020;10(11):1218\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMassey CJ, Ramos L, Beswick DM, Ramakrishnan VR, Humphries SM. Clinical Validation and Extension of an Automated, Deep Learning-Based Algorithm for Quantitative Sinus CT Analysis. AJNR Am J Neuroradiol. 2022;43(9):1318\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZou C, Ji H, Cui J, Qian B, Chen YC, Zhang Q, He S, Sui Y, Bai Y, Zhong Y, Zhang X, Ni T, Che Z. Preliminary study on AI-assisted diagnosis of bone remodeling in chronic maxillary sinusitis. BMC Med Imaging. 2024;24(1):140.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIsensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRonneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015: 18th international conference, Munich, Germany, October 5\u0026ndash;9, 2015, proceedings, part III 18. Springer international publishing, 2015: 234\u0026ndash;241.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Dai J, Wen X, Wang Y, Zhang Y, Wang N. Imaging and anatomical features of ethmomaxillary sinus and its differentiation from surrounding air cells. Surg Radiol Anat. 2018;40(2):207\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun R, Zhang M, Yang L, Yang S, Li N, Huang Y, Song H, Wang B, Huang C, Hou F, Wang H. Preoperative CT-based deep learning radiomics model to predict lymph node metastasis and patient prognosis in bladder cancer: a two-center study. Insights Imaging. 2024;15(1):21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Zhao J, Li Z, Yang M, Ye Z. Preoperative prediction of renal fibrous capsule invasion in clear cell renal cell carcinoma using CT-based radiomics model. Br J Radiol. 2024;97(1161):1557\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi X, Zhang J, Leng Y, Liu J, Li L, Wan T, Dong W, Fan B, Gong L. Preoperative prediction of histopathological grading in patients with chondrosarcoma using MRI-based radiomics with semantic features. BMC Med Imaging. 2024;24(1):171.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe S, Chen W, Wang X, Xie X, Liu F, Ma X, Li X, Li A, Feng X. Deep learning radiomics-based preoperative prediction of recurrence in chronic rhinosinusitis. iScience. 2023;26(4):106527.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLai S, Kang W, Chen Y, Zou J, Wang S, Zhang X, Zhang X, Lin Y. An End-to-End CRSwNP Prediction with Multichannel ResNet on Computed Tomography. Int J Biomed Imaging. 2024;2024:4960630.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"The infraorbital posterior ethmoid cells, Artificial intelligence, Deep learning, Efficient detection","lastPublishedDoi":"10.21203/rs.3.rs-6612929/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6612929/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjective\u003c/strong\u003e: To construct an AI-assisted model for identifying the infraorbital posterior ethmoid cells (IPECs) based on deep learning using sagittal CT images.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: Sagittal CT images of 277 samples with and 142 samples without IPECs were retrospectively collected. An experienced radiologist engaged in the relevant aspects picked a sagittal CT image that best showed IPECs. The images were randomly assigned to the training and test sets, with 541 sides in the training set and 97 sides in the test set. The training set was used to perform a five-fold cross-validation, and the results of each fold were used to predict the test set. The model was built using nnUNet, and its performance was evaluated using Dice and standard classification metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: The model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the additional set. Precision was 0.965 for the training set and 1.000 for the additional set, while sensitivity was 0.981 and 0.967, respectively. A comparison of the diagnostic efficacy between manual outlining by a less-experienced radiologist and AI-assisted outlining showed a significant improvement in detection efficiency (P \u0026lt; 0.05). The AI model aided correctly in identifying and outlining all IPECs, including 12 sides that the radiologist should improve portraying.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e: Artificial intelligence (AI) models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions.\u003c/p\u003e","manuscriptTitle":"Establishment of AI-assisted diagnosis of the infraorbital posterior ethmoid cells based on deep learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-16 12:04:07","doi":"10.21203/rs.3.rs-6612929/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-06-16T06:56:12+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-13T04:18:54+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-09T09:04:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-03T22:10:03+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"69053532399164848633261511641657848620","date":"2025-05-26T13:22:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"127966576475342812425625117983304555436","date":"2025-05-25T02:42:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"41096984621906209699071516101377046856","date":"2025-05-14T06:41:05+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-13T12:14:15+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-13T12:07:13+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-05-12T05:38:56+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-10T09:52:33+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-05-10T09:51:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d42a4618-4ffb-4bdb-8d63-6b88e8065fc4","owner":[],"postedDate":"May 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-07-28T16:05:59+00:00","versionOfRecord":{"articleIdentity":"rs-6612929","link":"https://doi.org/10.1186/s12880-025-01831-w","journal":{"identity":"bmc-medical-imaging","isVorOnly":false,"title":"BMC Medical Imaging"},"publishedOn":"2025-07-21 15:57:58","publishedOnDateReadable":"July 21st, 2025"},"versionCreatedAt":"2025-05-16 12:04:07","video":"","vorDoi":"10.1186/s12880-025-01831-w","vorDoiUrl":"https://doi.org/10.1186/s12880-025-01831-w","workflowStages":[]},"version":"v1","identity":"rs-6612929","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6612929","identity":"rs-6612929","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}