Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images

Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images | 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 Article Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images Jaepil Ko, MinHye Kang, Young Joon Jun This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5221450/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 08 Jul, 2025 Read the published version in Scientific Reports → Version 1 posted 8 You are reading this latest preprint version Abstract Purpose Allergic rhinitis typically has edematous and pale turbinates or erythematous and inflamed turbinates. While traditional approaches include using skin prick tests (SPT) to determine the presence of AR, It is often not related to actual symptoms, and it is an invasive test. We use deep learning to analyze nasal endoscopy images to investigate a quantitative method for diagnosing allergic rhinitis. Methods Traditional machine learning-based diagnostic techniques have relied on structured clinical datasets featuring statistical data such as demographic characteristics, symptom severity, and clinical test results. In contrast, we propose a novel approach to use endoscopy image data to analyze the color distribution in the inferior turbinate region of patients with allergic rhinitis using the CIE-Lab color space and extract the adaptive histogram features that are used to explore and find suitable feature extraction methods and deep learning model architectures. Results Our proposed model achieves a promising diagnostic accuracy of 90.80% for images exhibiting AR symptoms. Future research will expand the dataset to include a broader spectrum of symptomatic and asymptomatic images to enhance model robustness and investigate the potential of optical analysis as a non-invasive diagnostic method for AR. Conclusion This study introduced a novel approach to diagnosing allergic rhinitis using nasal endoscopy images. Our approach analyzed the color distribution of the inferior turbinates within the LAB color space, extracted important features from endoscopy images using both CNN feature extraction and histograms, and performed classification through SVM and fully connected classifiers. Biological sciences/Computational biology and bioinformatics/Image processing Biological sciences/Computational biology and bioinformatics/Computational models Health sciences/Health care/Diagnosis/Physical examination Health sciences/Medical research/Experimental models of disease Allergic Rhinitis Diagnosis Nasal Endoscope SVM CNN Classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 6 Figure 7 Figure 8 Introduction Allergic rhinitis (AR) is an IgE-mediated inflammatory response in the nasal mucosa following exposure to certain antigens. These triggers sneezing, nasal congestion, and runny nose ( 1 ). Recently, the incidents of allergic rhinitis have been increasing due to increased concentrations of PM 10 and 2.5 dust caused by air pollution ( 2 ). According to the Korea National Health Insurance Service, allergic rhinitis patients and related medical expenses constantly increase, resulting in a significant burden ( 3 ). Diagnosing allergic rhinitis typically involves taking a patient's medical history, skin prick tests, and measuring specific IgE antibodies. Taking a medical history is straightforward and cost-effective but can be imprecise due to its reliance on subjective symptoms and family health background. Skin prick tests, while quick [4], may not be suitable for all patients due to age, skin conditions, or certain medications ( 5 ). Furthermore, specific IgE antibody tests are not limited by the subject's condition [6] but are expensive and time-consuming ( 4 , 6 ). Skin prick tests and specific IgE tests can analyze the causative antigens of allergic rhinitis and are qualitative diagnostic methods, but they are invasive and expensive. Moreover, in the case of objective measurement methods, the relationship with the patient's symptoms is still ambiguous. Consequently, the diagnoses that rely on patient history are subjective and need more precision and objectivity for accurate diagnosis. Guidelines for diagnosing allergic rhinitis include nasal examinations ( 7 ). A typical sign of allergic rhinitis, observed through nasal endoscopy, is the appearance of pale and swollen inferior turbinates ( 8 ). Chronic edema and pale mucosa may occur in patients who have suffered long-term AR. This distinguishment is not a quantitative diagnostic value. Studies using RGB color analysis have found that these turbinates in affected patients display increased levels of green and blue hues ( 9 ). Additionally, HSL color analysis indicates that the turbinates are paler and smoother than those of non-affected individuals ( 10 ). Despite various diagnostic approaches, allergic rhinitis diagnosis remains challenging due to the subjective nature of clinical histories, invasiveness and limitations of skin prick and IgE antibody tests, and the qualitative nature of current nasal endoscopy evaluations. Thus, a substantial research gap exists in developing standardized, non-invasive, and quantitative diagnostic methods for allergic rhinitis that remain cost-effective and efficient. This study aims to investigate AR diagnostic devices, measure and quantify the nasal endoscopy image via optical analysis, and compare it to those in AR patients. Our approach involves two main stages: feature extraction and classification. Initially, we analyze the color distribution of the inferior turbinates using the Lab color space, comparing allergic rhinitis patients to non-affected individuals. To understand the endoscopy images and effectively recognize turbinate, we extract meaningful feature vectors in two ways. One uses a histogram, and the second uses a pre-trained convolutional neural network (CNN). These features are then evaluated utilizing support vector machines (SVM) and fully connected layers (FC) as classifiers. Methods 1. Acquisition of Nasal Endoscopy Images The data used in this study was approved by the Institutional Review Board of Gumi Hospital, Sunchunhyang University. Images of the left and right nasal passages were collected who underwent nasal endoscopy between March 2019 and March 2020. Written informed consent was received from all patients. The subjects were then subjected to the Multiple Allergen Simultaneous Test (MAST), a specific IgE antibody test, to collect information on the presence of allergic rhinitis. Exclusion criteria were patients who had undergone surgery, had a history of sinusitis, or had a nasal tumor. To keep the images consistent, all images were taken with the same white balance adjustment. All images were acquired with the same white balance settings before acquisition and adjusted to equalize the endoscope distance to minimize image distortion. Only images from a single physician were analyzed to reduce inter-physician variation. The nasal endoscopic images have a resolution of 640×480 and a circular view of the nasal endoscopic area on a black background. The left and right nasal cavities are vertically symmetrical and include the nasal septum and the inferior turbinate (IT) in one image. The nasal septum is the middle membrane that divides the two nasal cavities and appears as a smooth wall on endonasal imaging, while the inferior turbinate is curved (Fig. 1 ). 2. Histogram Features from the Inferior Turbinate Area Using the Lab Color Space RGB colors have the disadvantage of being inconsistent and can be affected by ambient light or environment. There are alternatives called HSL color space or Lab color space. The International Commission on Luminance prescribes lab color space, where L represents brightness, a represents the degree of reddish green, and b represents the degree of yellowish blue. Lab color model can be consistent regardless of monitor changes or printer color variations. Similar to the human visual system. This study compares the color distribution of the inferior turbinates of normal and allergic rhinitis groups in the Lab color space. To this end, we implemented a simple s/w to extract histograms of the inferior turbinate region, as shown in Fig. 2 . This s/w consists of four main steps (Fig. 3 ). First, we set up a grid of 6x5 cells on the inferior turbinate area. Then, we exclude cells that do not contain the inferior turbinates. Since the structure of the nasal cavity is different for each person and the point at which the nasal endoscope is taken is different, the shape and area of the inferior turbinates shown in each image are very different. As a result, the inferior turbinate region is often not included within the 30 cells of the grid created initially in the first step. Therefore, we manually specify and exclude cells that do not contain the inferior turbinate region. We then calculate the average pixel value of the Lab channel within the cell. Finally, we compute a histogram from the average values. The histogram is computed for the a and b channels representing colors only. We carefully determined that the number of histogram bins should be 5, with the range of each bin as follows: 5 or less, 6 to 10, 11 to 15, 16 to 20, and 21 or more for the channel a, and − 5 or less, -4 to 0, 1 to 5, 6 to 10, and 11 or more for the channel b. 3. CNN Features from the Inferior Turbinate Area Using the RGB Color Space We extract CNN features for better representation. We employ the Inception v3 ( 5 ) model that has been pre-trained on ImageNet weights. We selected Inception v3 pre-trained on ImageNet due to its efficient handling of multi-scale features. Given the variability in the nasal endoscopy images (scale variations due to differing distances between the camera and turbinate regions), Inception v3’s multi-kernel architecture effectively captures features at multiple scales, thus making it ideal for our study. Considering that the Inception model was originally trained on RGB color images, we opted to use the RGB color model for consistency. Figure 4 illustrates the process of acquiring image patches for the model inputs. First, we apply a median filter to extract features focusing on the color information of the inferior turbinates (as shown in (a) in Fig. 4 ). We use a median filter as a blurring technique because nasal endoscopy images often contain foreign bodies, such as runny nose secretions and light reflections from endoscope illumination. Notably, the blurring does not result in the loss of information since we use color information. The filter size was chosen to be 21, a value determined through simple experimentation, as it is sufficiently larger than the block size. Subsequently, the inferior turbinate region is labeled with polygons to crop patches for the Inception model. Next, we collect 100 patches of size 32×32 from random locations within the labeled region ((b) in Fig. 4 ). The patches we obtained are normalized to have values between 0 and 1 by dividing by the maximum value of the color space range, and passed through the Inception model to extract 2048-dimensional features. This feature extraction method is simpler than histogram-based methods as it only involves basic polygon labeling without requiring meticulous bin and range settings. 4. Classification Models In this paper, we analyze the binary classification performance of allergic rhinitis's presence or absence using SVMs ( 12 ) and fully connected classifiers. We use the SVM classifier with the Radial Basis Function (RBF) ( 13 ) kernel. For our fully connected classifier, we designed a simple fully connected network, as shown in Fig. 5 , leveraging a pre-trained model for feature extraction. It comprises a fully connected layer with dropout ( 14 ) for regularization, followed by a softmax layer for the final classification. It includes two dense layers and a dropout layer, a widely used regularization technique, forming a typical and basic classifier design. Results 1. Nasal Endoscopy Image Dataset After conducting endoscopic imaging on 150 patients, we selected good-quality images from 46 patients. 18 images were taken from 18 individuals in the normal group, and 74 images were taken from 28 individuals in the allergic rhinitis group, totaling 92 images. Figure 6 presents the images obtained for the dataset. Notably, our data set has a significant data imbalance problem. To address this issue, we employed meticulous and thoughtful techniques during image preprocessing and analysis to ensure the reliability and validity of our findings despite the imbalance. The varying imaging distances caused scale differences in nasal cavity images, which the Inception model effectively handled using multiple kernel sizes. 2. Comparison Result of Lab Color Distributions between Normal and Allergic Rhinitis Groups We compared the color distributions in the Lab color space between the normal and allergic rhinitis groups using 92 images, as detailed in the Methods section. Figure 7 clearly demonstrates the differences between these two distributions. In channel a, the normal group exhibits a higher proportion of histogram values of 16 or higher compared to the allergic rhinitis group. Conversely, in channel b, the allergic rhinitis group shows a higher proportion of histogram values of 0 or lower than the normal group. These observations confirm that the color distribution in the inferior turbinates differs between the normal and allergic rhinitis groups. 3. Experimental Strategies for Addressing Class Imbalance Problem As previously mentioned, the dataset suffers from severe class imbalance. To address this issue, we implemented three strategies in our experiments. Firstly, we repeated experiments by the following cross-validation: randomly dividing the data in ratios of 7:3, 6:4, and 5:5 to create training and validation datasets. The proportion and class weight of data for each split is detailed in Table 1 . Secondly, we conducted additional experiments employing class weights. For each class weight \(\:{\text{w}}_{i}\) , we calculate as follows: $$\:{w}_{i}=\frac{T}{C\times\:{C}_{i}}$$ 1 where \(\:\text{T}\) is the total number of images, \(\:\text{C}\) is the total number of classes, and \(\:{C}_{i}\) is the number of images in the \(\:i\) -th class on training data set. Finally, we adapted F1-score as well as accuracy as performance metrics. F1-score is a statistical measure for a binary classification model when the class distribution is uneven. Table 1 Number of images and class weight per class by data split ratio. Data Class 7:3 6:4 5:5 Train (weight) Normal class 12 (2.63) 10 (2.70) 9 (2.55) Mast class 51 (0.62) 44 (0.61) 37 (0.62) Test Normal class 6 8 9 Mast class 23 30 37 Experiments with the SVM classifier are conducted with parameter C set at 1, 10, 100, 1000, and 10000, and gamma at 0.1, 0.01, 0.001, 0.0001, and 0.00001. For the fully connected classifier, we use a learning rate of 0.00001, cross-entropy as the loss function, and Adam optimization as the optimizer. 4. Experimental Results on Classification First, we compared the performance of histogram features and CNN features using the SVM classifier. Tables 2 and 3 represent the experimental results for the SVM classifier applied to histogram features and CNN features, respectively. Additionally, Table 2 includes the performance across three different color spaces. For each image, we extracted 100 patches and performed classification individually on each patch. Final classification results per image were obtained through majority voting among the patches, ensuring robust classification by aggregating patch-level predictions. Table 2 Performance of histogram features in ab color space using SVM classifier. Data split ratio 7:3 6:4 5:5 Class weight w/o w w/o w w/o w Accuracy 89.66 89.66 89.47 89.47 89.13 86.96 F1-score 0.9388 0.9362 0.9375 0.9355 0.9351 0.9167 Table 3 Performance of CNN features in RGB, Lab, and ab color spaces using SVM classifier. Data split ratio 7:3 6:4 5:5 Class weight w/o w w/o w w/o w RGB Accuracy 88.21 88.69 86.71 86.34 87.96 84.63 F1-score 0.928 0.9286 0.9188 0.9141 0.9267 0.9233 Lab Accuracy 84.45 83.76 81.71 81.39 82.91 82.87 F1-score 0.9077 0.9052 0.8915 0.8909 0.8987 0.8986 Ab Accuracy 82.24 81.72 82.21 81.82 84.72 84.26 F1-score 0.8791 0.8926 0.8952 0.8924 0.9113 0.9072 Tables 2 and 3 demonstrate the improvement in performance as the proportion of training data increases. Notably, both tables indicate that the application of class weights had no significant effect. Furthermore, Histogram-based features consistently outperformed CNN-extracted features when classified by SVM, likely due to their direct and targeted representation of color distribution differences, which are critical in identifying allergic rhinitis conditions from limited training samples. All the models with class weights showed a slight decrease in performance except for the model using CNN features in the 7:3 split on RGB color space. Regarding color spaces, we can compare the ab color space, which solely contains color components, against Lab and RGB, both of which include color and brightness components. Table 3 shows that it is difficult to determine the superiority of ab or Lab; however, the RGB color space was clearly superior to both spaces in the case of CNN features. Table 2 was produced by selecting the highest values from those obtained by varying the SVM parameters, as shown in Fig. 8 . The highest accuracy of 89.66% and an F1-score of 0.9388 were achieved with C = 100 and gamma = 0.001. Secondly, we compared the performance of CNN features across three color spaces using fully connected networks, with data divided in ratios of 7:3, 6:4, and 5:5, as shown in Table 4 . Multiple data splits were employed to ensure robustness and consistency of results under varying training and testing conditions. Table 4 Performance of CNN features in RGB, Lab, and ab color spaces using fully connected networks. Data split ratio 7:3 6:4 5:5 Class weight w/o w w/o w w/o w RGB Accuracy 86.2 93.1 84.21 93.1 80.43 86.2 F1-score 0.92 0.9545 0.909 0.9545 0.9 0.9047 Lab Accuracy 86.2 92.13 82.21 92.1 80.43 81.57 F1-score 0.92 0.9475 0.8823 0.9473 0.8915 0.8771 Ab Accuracy 89.65 89.13 83.14 89.13 80.43 84.78 F1-score 0.9387 0.9367 0.8823 0.9295 0.8915 0.9041 The highest accuracy of 93.1% and F1-score of 0.9545 were significantly better than those achieved with the SVM classifier. Unlike the SVM case, we observed significant performance improvements with class weights in all cases but one case. This outcome demonstrates that class weights in neural networks effectively address data imbalance problems. Similar to previous experiments, the RGB color space clearly exhibited the highest performance among the color spaces. Discussions Allergic rhinitis, a chronic condition that significantly impacts a patient's quality of life, is currently diagnosed using either non-invasive or invasive methods, both of which are time-consuming and costly. This paper presents a novel approach to a deep learning-based allergic rhinitis diagnosis model that utilizes nasal endoscopic images. This method offers a less time-consuming, cost-effective, non-invasive, and qualitative alternative to the existing diagnostic methods. Allergic rhinitis affects 10–20% of the population, with 4–8% annual growth rates. Despite the surge in patent applications since 2015, advancements in diagnostic technologies have been relatively weak (especially in the AR and nasal obstruction diagnostic/decision category, outsourced patent research, and patent attorneys). HSV color spaces are alternative representations of the RGB (red, green, blue) color model, designed in the 1970s by computer graphics researchers to more closely align with how human vision perceives color-making attributes ( 15 ). The colors of each hue are arranged in a radial slice, saturation consists of dimensions resembling various tints of brightly colored paint, and value mentions dimensions resembling the mixture of those paints with varying amounts of black or white paint. Lab color spaces can be closer to human vision than HSV and consistent regardless of monitor changes or printer color variations. By analyzing the inferior turbinate color distribution in the Lab color space, we found that the inferior turbinate color distribution of the normal and allergic rhinitis groups differed. Based on this, we extracted CNN features by extracting histogram features and passing them through a pre-trained Inception v3 model on ImageNet weights ( 5 , 11 ). We conducted comparative experiments using SVM and fully connected classifiers to analyze which feature extraction method is more efficient and which learning technique performs better. SVM is widely regarded as an effective binary classifier. SVM classifiers have the advantage of effectively determining optimal decision boundaries even when sample sizes are limited, thus ensuring reliable generalization from limited training data. Accordingly, we selected SVM as the representative machine learning classifier. The experimental results show that the best accuracy of histogram features with the SVM classifier is 0.8966 and F1-score is 0.9388, while the best accuracy of CNN features with the SVM classifier is 0.8821 and F1-score is 0.928, which shows that histogram features perform better. Using a fully connected classifier on the CNN features achieved the best results, with the best accuracy of 0.9310 and F1-score of 0.9545 ( 11 , 12 ). Although the histogram features performed better than those extracted from the CNN trained for general purposes, they require manual effort in the inferior turbinate detection process. Although CNN features perform worse than histogram features, we found that better performance can be achieved by applying a fully connected classifier with class weights ( 11 ). It is important to acknowledge the limitations of this study, which include factors such as the nasal endoscopy angle, color adjustment, and analysis program. These limitations highlight the need for continued research and innovation in these areas, underscoring the potential for further advancements in allergic rhinitis diagnosis and treatment. We adopted an alternative approach to minimize this bias as much as possible, conducting multiple experiments with different random split ratios (e.g., 7:3, 6:4, 5:5). We ensured patient-wise separation to minimize overlap. Although this is not a traditional n-fold cross-validation method, it represents the most appropriate strategy given the characteristics of our dataset. Furthermore, for this study, we demonstrate that our deep learning-based feature extraction and classifier architecture outperforms conventional hand-crafted features and machine learning-based classifiers. Although this is not a traditional n-fold cross-validation method, it represents the most appropriate strategy given the characteristics of our dataset. Furthermore, this study demonstrated that our deep learning-based feature extraction and classifier architecture outperforms conventional hand-crafted features and machine learning-based classifiers. In general, when the amount of data is limited, confidence intervals are appropriate. However, in the case of neural networks, it is common not to report them when the number of test samples is sufficiently large. In our CNN + FC feature extraction and classification approach, we extract 100 additional patches per image (as described below in Fig. 4 ). Therefore, testing on just 10 images effectively results in classification being performed on 1,000 patches. The final performance is then evaluated by aggregating the classification results of these patches using a voting mechanism. In other words, since the classifier operates at the patch level, the number of test samples is sufficiently large to evaluate the model’s performance reliably. In addition, our results rely primarily on accuracy and F1-score due to our initial experimental scope. We chose accuracy as a clear, easily interpretable measure of overall model performance and F1-score as it effectively balances precision and recall, providing valuable insight, particularly for imbalanced datasets such as ours. Nasal endoscopy is likely to vary slightly from doctor to doctor, which can lead to slight angular variations. A certain depth or distance is important because the amount of light can vary. To reduce these errors, a single ENT surgeon performed the study, and white balance and brightness adjustments were performed under the same conditions before the images were acquired. The image was also acquired with a constant depth control to control the amount of light entering the image. Nevertheless, some issues will inevitably lead to differences, which need further study. One solution is the recent development of artificial intelligence in image adjustment, which can be used to adjust the angle slightly. Additional research is needed. Image analysis, particularly in medical diagnostics, is advancing rapidly. Our study utilized SVM alongside various variables but explored other image analysis methods, including VGG, Inception, Xception, ResNet, DenseNet, and IncRes. The images collected in this study were not captured from a fixed distance. Instead, the distance to the target varied depending on the shape of the patient’s nasal cavity and the conditions at the time of capture. This led to scale variations of the nasal cavity within the images. Inception, characterized by its use of multiple kernel sizes, is particularly effective in handling such scale variations of target objects. Various models, such as VGG, Xception, ResNet, and Inception, can be considered for CNN-based feature extraction in transfer learning. These models differ in usability, the number of parameters, and the characteristics of the extracted features, and their performance may vary depending on the specific image classification task. In this study, we used features extracted directly from a pre-trained CNN rather than fine-tuning the CNN with our dataset. We aimed to demonstrate that general-purpose image features learned from ImageNet could outperform even carefully crafted histogram-based features. Therefore, the objective of this study was to obtain general image features. Despite these innovations, our findings indicated that newer pre-trained models could have yielded better performance and produced consistent results. For future investigations into mucous membrane analysis, such as those conducted in this study, alternative models might be more effective, or there may be a need to develop new pre-trained models specifically tailored for this application. In summary, as demonstrated in this study, the optical analysis of nasal endoscopy images could be a valuable adjunct to non-invasive measurement methods for allergic rhinitis. The potential benefits of this new method include improved efficiency, reduced costs, and enhanced patient comfort. However, further studies are necessary to fully evaluate and validate these new methods. Conclusion This study introduced a novel approach to diagnosing allergic rhinitis using nasal endoscopy images. Our approach analyzed the color distribution of the inferior turbinates within the LAB color space, extracted important features from endoscopy images using both CNN feature extraction and histograms, and performed classification through SVM and fully connected classifiers. Our findings indicated that while histogram features combined with SVM classifiers showed high accuracy and F1 scores, the best results were obtained using CNN features with a fully connected classifier, achieving 90.8% diagnostic accuracy. This suggests that deep learning frameworks can enhance diagnostic accuracy and efficiency when properly tuned and applied to specific medical imaging tasks. However, the study also recognized limitations due to the inherent variability in nasal endoscopy procedures, such as differences in angle and lighting conditions, which can affect image analysis. Future work will address these challenges by refining image capture consistency and exploring the use of advanced image processing technologies. Abbreviations AR (Allergic rhinitis), SVM (support vector machines), CNN (convolutional neural network), SPT (skin prick tests), MAST (Multiple Allergen Simultaneous Test), IT (inferior turbinate), RBF (Radial Basis Function), RGB (red, green, blue), HSV (hue, saturation, value), LAB (CIELAB), FC (fully connected layers), VGG (Visual Geometry Group Declarations The authors declare no conflict of interest Data Availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. E-mail: [email protected] (corresponding author) Acknowledgments Author Contributions Conceptualization: Young Joon Jun Data curation: MinHye Kang, JaePil Ko, Young Joon Jun Funding acquisition: Young Joon Jun Methodology—clinical: Young Joon Jun Methodology—computing: MinHye Kang, JaePil Ko Project administration: JaePil Ko, Young Joon Jun. Visualization: MinHye Kang, JaePil Ko, Young Joon Jun Writing—original draft: MinHye Kang, Young Joon Jun Writing—review & editing: JaePil Ko, Young Joon Jun Funding : This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (2019R1I1A3A01063980) Ethics committee approval and subject consent The study was approved by SoonChunHyang University Gumi Hospital and Uijeongbu Eulji University Hospital Institutional Review Board (IRB). Patients received informed consent. All research has been performed in accordance with the Declaration of Helsinki. References Pawankar R, Bunnag C, Khaltaev N, Bousquet J. Allergic Rhinitis and Its Impact on Asthma in Asia Pacific and the ARIA Update 2008. World Allergy Organ J. 2012;5(Suppl 3):S212-7. Lee K-I, Chung Y-J, Mo J-H. The impact of air pollution on allergic rhinitis. Allergy, Asthma & Respiratory Disease. 2021;9(1). Yoo KH, Ahn HR, Park JK, Kim JW, Nam GH, Hong SK, et al. Burden of Respiratory Disease in Korea: An Observational Study on Allergic Rhinitis, Asthma, COPD, and Rhinosinusitis. Allergy Asthma Immunol Res. 2016;8(6):527-34. Yang SI, Lee IH, Kim M, Ryu G, Kang SY, Kim MA, et al. KAAACI Allergic Rhinitis Guidelines: Part 1. Update in Pharmacotherapy. Allergy Asthma Immunol Res. 2023;15(1):19-31. Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the Inception Architecture for Computer Vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)2016. p. 2818-26. Kim YH, Kim K-S. Diagnosis and treatment of allergic rhinitis. Journal of the Korean Medical Association. 2010;53(9). Brozek JL, Bousquet J, Agache I, Agarwal A, Bachert C, Bosnic-Anticevich S, et al. Allergic Rhinitis and its Impact on Asthma (ARIA) guidelines-2016 revision. J Allergy Clin Immunol. 2017;140(4):950-8. Seidman MD, Gurgel RK, Lin SY, Schwartz SR, Baroody FM, Bonner JR, et al. Clinical practice guideline: Allergic rhinitis. Otolaryngol Head Neck Surg. 2015;152(1 Suppl):S1-43. Joko H, Hyodo M, Gyo K, Yumoto E. Chromametric assessment of nasal mucosal color and its application in patients with nasal allergy. Am J Rhinol. 2002;16(1):11-6. Bae S, Jun YJ. Optical Analysis of Nasal Endoscopic Images From a Patient With Severe Acute Respiratory Syndrome Coronavirus 2. Journal of Rhinology. 2022;29(2):96-100. O’Shea K, Nash R. An Introduction to Convolutional Neural Networks. ArXiv. 2015;abs/1511.08458. Scholkopf B, Smola AJ. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond: MIT Press; 2001. Machine Learning and Its Applications2001. Srivastava N, Hinton GE, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014;15:1929-58. Ibraheem NA, Hasan MM, Khan RZ, Mishra PK. Understanding color models: a review. ARPN Journal of science and technology. 2012;2(3):265-75. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 08 Jul, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 27 May, 2025 Reviews received at journal 22 May, 2025 Reviewers agreed at journal 12 May, 2025 Reviews received at journal 25 Apr, 2025 Reviewers agreed at journal 08 Apr, 2025 Reviewers invited by journal 08 Apr, 2025 Submission checks completed at journal 07 Apr, 2025 First submitted to journal 25 Mar, 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-5221450","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":439935146,"identity":"983c4f6d-c270-439b-8e6c-de511848b6c9","order_by":0,"name":"Jaepil Ko","email":"","orcid":"","institution":"Kumoh National Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Jaepil","middleName":"","lastName":"Ko","suffix":""},{"id":439935148,"identity":"bcb7d2fa-9c47-4b18-accd-2b51b7dc698a","order_by":1,"name":"MinHye Kang","email":"","orcid":"","institution":"Kumoh National Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"MinHye","middleName":"","lastName":"Kang","suffix":""},{"id":439935149,"identity":"1a7faf85-663d-40a0-8945-ea908aed69c2","order_by":2,"name":"Young Joon Jun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxklEQVRIiWNgGAWjYLACCQYGORB94AEpWozBWhJIsSixAUQSpUV+Ru6zD5Ztd9Lnhx1+CLTFTk63gYAWgxvpxjMk257lbrydZgDUkmxsdoCQFok0ZgbJtsO5G2cngLQcSNxGSIv8DIiWdMPZ6R+I08JwA6IlQV46h0hbDM48Y2aQOHfYcIN0TsGBBAMi/CLfnsbMLFF2WF5+dvrmDx8q7OQIamEQSGBglgBZB1ZpQEg5CPAfYGD8ALKugRjVo2AUjIJRMCIBAMH+Q+2+KGwgAAAAAElFTkSuQmCC","orcid":"","institution":"Uijeongbu Eulji Medical Center, Eulji University","correspondingAuthor":true,"prefix":"","firstName":"Young","middleName":"Joon","lastName":"Jun","suffix":""}],"badges":[],"createdAt":"2024-10-08 03:08:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5221450/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5221450/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-10087-x","type":"published","date":"2025-07-08T15:57:53+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":80275738,"identity":"29ea88d5-e1e5-40fd-bed8-6a5c5a219e0a","added_by":"auto","created_at":"2025-04-10 04:55:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":113960,"visible":true,"origin":"","legend":"\u003cp\u003eNasal structures of nasal endoscopy image.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/4dca9de29f6d25ea244473c6.png"},{"id":80275740,"identity":"05009848-86d7-4ff1-baa0-3ba071117c59","added_by":"auto","created_at":"2025-04-10 04:55:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":115050,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of Lab histogram analysis s/w.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/fad3db7580814e3305c6eeb9.png"},{"id":80275741,"identity":"f5321185-cfb3-4314-9497-739a9923a412","added_by":"auto","created_at":"2025-04-10 04:55:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":30949,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of extracting the histogram in the Lab color model.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/c639e272991b30624da72269.png"},{"id":80276184,"identity":"cb6e6e1a-7f1e-440b-8166-8c70d7bfe908","added_by":"auto","created_at":"2025-04-10 05:03:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":48891,"visible":true,"origin":"","legend":"\u003cp\u003eInferior turbinate area in green and random box in red for patch and (b) its patch images.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/8b7d95d30a559fc7aafc49ce.png"},{"id":80275768,"identity":"ca9b453d-67c0-401d-a52e-425d45c42b98","added_by":"auto","created_at":"2025-04-10 04:55:16","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":113570,"visible":true,"origin":"","legend":"\u003cp\u003eNasal endoscopy images of our data set.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/4050b37a779778e14d67f3b0.png"},{"id":80276186,"identity":"259f6c8b-4810-4aa8-a6de-dfcac55b72f8","added_by":"auto","created_at":"2025-04-10 05:03:15","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":100349,"visible":true,"origin":"","legend":"\u003cp\u003eColor distribution of normal and allergic rhinitis groups in Lab color space.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/20f530450acaf7be7039d0ce.png"},{"id":80275762,"identity":"4e80510b-46f6-4d5a-bfb7-822379a3229c","added_by":"auto","created_at":"2025-04-10 04:55:16","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":88808,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmaps on test accuracy and F1-score by the SVM parameters C and gamma for histogram features in the 7:3 split without applying class weight. The boxes highlighted in red indicate the highest values.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/3a9b9858a3f43ae3b7b08587.png"},{"id":86700365,"identity":"77fc781f-dd6a-4492-88c5-c99da7591597","added_by":"auto","created_at":"2025-07-14 16:12:07","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1326409,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5221450/v1/e013da85-19ad-4bb0-b9b1-72056b3ec8a6.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAllergic rhinitis (AR) is an IgE-mediated inflammatory response in the nasal mucosa following exposure to certain antigens. These triggers sneezing, nasal congestion, and runny nose (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Recently, the incidents of allergic rhinitis have been increasing due to increased concentrations of PM 10 and 2.5 dust caused by air pollution (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). According to the Korea National Health Insurance Service, allergic rhinitis patients and related medical expenses constantly increase, resulting in a significant burden (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDiagnosing allergic rhinitis typically involves taking a patient's medical history, skin prick tests, and measuring specific IgE antibodies. Taking a medical history is straightforward and cost-effective but can be imprecise due to its reliance on subjective symptoms and family health background. Skin prick tests, while quick [4], may not be suitable for all patients due to age, skin conditions, or certain medications (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Furthermore, specific IgE antibody tests are not limited by the subject's condition [6] but are expensive and time-consuming (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). Skin prick tests and specific IgE tests can analyze the causative antigens of allergic rhinitis and are qualitative diagnostic methods, but they are invasive and expensive. Moreover, in the case of objective measurement methods, the relationship with the patient's symptoms is still ambiguous. Consequently, the diagnoses that rely on patient history are subjective and need more precision and objectivity for accurate diagnosis.\u003c/p\u003e \u003cp\u003eGuidelines for diagnosing allergic rhinitis include nasal examinations (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). A typical sign of allergic rhinitis, observed through nasal endoscopy, is the appearance of pale and swollen inferior turbinates (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Chronic edema and pale mucosa may occur in patients who have suffered long-term AR. This distinguishment is not a quantitative diagnostic value.\u003c/p\u003e \u003cp\u003eStudies using RGB color analysis have found that these turbinates in affected patients display increased levels of green and blue hues (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). Additionally, HSL color analysis indicates that the turbinates are paler and smoother than those of non-affected individuals (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Despite various diagnostic approaches, allergic rhinitis diagnosis remains challenging due to the subjective nature of clinical histories, invasiveness and limitations of skin prick and IgE antibody tests, and the qualitative nature of current nasal endoscopy evaluations. Thus, a substantial research gap exists in developing standardized, non-invasive, and quantitative diagnostic methods for allergic rhinitis that remain cost-effective and efficient.\u003c/p\u003e \u003cp\u003eThis study aims to investigate AR diagnostic devices, measure and quantify the nasal endoscopy image via optical analysis, and compare it to those in AR patients. Our approach involves two main stages: feature extraction and classification. Initially, we analyze the color distribution of the inferior turbinates using the Lab color space, comparing allergic rhinitis patients to non-affected individuals. To understand the endoscopy images and effectively recognize turbinate, we extract meaningful feature vectors in two ways. One uses a histogram, and the second uses a pre-trained convolutional neural network (CNN). These features are then evaluated utilizing support vector machines (SVM) and fully connected layers (FC) as classifiers.\u003c/p\u003e"},{"header":"Methods","content":"\n\u003ch3\u003e1. Acquisition of Nasal Endoscopy Images\u003c/h3\u003e\n\u003cp\u003e The data used in this study was approved by the Institutional Review Board of Gumi Hospital, Sunchunhyang University. Images of the left and right nasal passages were collected who underwent nasal endoscopy between March 2019 and March 2020. Written informed consent was received from all patients. The subjects were then subjected to the Multiple Allergen Simultaneous Test (MAST), a specific IgE antibody test, to collect information on the presence of allergic rhinitis. Exclusion criteria were patients who had undergone surgery, had a history of sinusitis, or had a nasal tumor.\u003c/p\u003e \u003cp\u003eTo keep the images consistent, all images were taken with the same white balance adjustment. All images were acquired with the same white balance settings before acquisition and adjusted to equalize the endoscope distance to minimize image distortion. Only images from a single physician were analyzed to reduce inter-physician variation.\u003c/p\u003e \u003cp\u003eThe nasal endoscopic images have a resolution of 640\u0026times;480 and a circular view of the nasal endoscopic area on a black background. The left and right nasal cavities are vertically symmetrical and include the nasal septum and the inferior turbinate (IT) in one image. The nasal septum is the middle membrane that divides the two nasal cavities and appears as a smooth wall on endonasal imaging, while the inferior turbinate is curved (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2. Histogram Features from the Inferior Turbinate Area Using the Lab Color Space\u003c/h2\u003e \u003cp\u003eRGB colors have the disadvantage of being inconsistent and can be affected by ambient light or environment. There are alternatives called HSL color space or Lab color space. The International Commission on Luminance prescribes lab color space, where L represents brightness, a represents the degree of reddish green, and b represents the degree of yellowish blue. Lab color model can be consistent regardless of monitor changes or printer color variations. Similar to the human visual system.\u003c/p\u003e \u003cp\u003eThis study compares the color distribution of the inferior turbinates of normal and allergic rhinitis groups in the Lab color space. To this end, we implemented a simple s/w to extract histograms of the inferior turbinate region, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis s/w consists of four main steps (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). First, we set up a grid of 6x5 cells on the inferior turbinate area. Then, we exclude cells that do not contain the inferior turbinates. Since the structure of the nasal cavity is different for each person and the point at which the nasal endoscope is taken is different, the shape and area of the inferior turbinates shown in each image are very different. As a result, the inferior turbinate region is often not included within the 30 cells of the grid created initially in the first step. Therefore, we manually specify and exclude cells that do not contain the inferior turbinate region. We then calculate the average pixel value of the Lab channel within the cell. Finally, we compute a histogram from the average values. The histogram is computed for the a and b channels representing colors only. We carefully determined that the number of histogram bins should be 5, with the range of each bin as follows: 5 or less, 6 to 10, 11 to 15, 16 to 20, and 21 or more for the channel a, and \u0026minus;\u0026thinsp;5 or less, -4 to 0, 1 to 5, 6 to 10, and 11 or more for the channel b.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e3. CNN Features from the Inferior Turbinate Area Using the RGB Color Space\u003c/h3\u003e\n\u003cp\u003eWe extract CNN features for better representation. We employ the Inception v3 (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) model that has been pre-trained on ImageNet weights. We selected Inception v3 pre-trained on ImageNet due to its efficient handling of multi-scale features. Given the variability in the nasal endoscopy images (scale variations due to differing distances between the camera and turbinate regions), Inception v3\u0026rsquo;s multi-kernel architecture effectively captures features at multiple scales, thus making it ideal for our study. Considering that the Inception model was originally trained on RGB color images, we opted to use the RGB color model for consistency. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the process of acquiring image patches for the model inputs.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFirst, we apply a median filter to extract features focusing on the color information of the inferior turbinates (as shown in (a) in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). We use a median filter as a blurring technique because nasal endoscopy images often contain foreign bodies, such as runny nose secretions and light reflections from endoscope illumination. Notably, the blurring does not result in the loss of information since we use color information. The filter size was chosen to be 21, a value determined through simple experimentation, as it is sufficiently larger than the block size. Subsequently, the inferior turbinate region is labeled with polygons to crop patches for the Inception model. Next, we collect 100 patches of size 32\u0026times;32 from random locations within the labeled region ((b) in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The patches we obtained are normalized to have values between 0 and 1 by dividing by the maximum value of the color space range, and passed through the Inception model to extract 2048-dimensional features.\u003c/p\u003e \u003cp\u003eThis feature extraction method is simpler than histogram-based methods as it only involves basic polygon labeling without requiring meticulous bin and range settings.\u003c/p\u003e\n\u003ch3\u003e4. Classification Models\u003c/h3\u003e\n\u003cp\u003eIn this paper, we analyze the binary classification performance of allergic rhinitis's presence or absence using SVMs (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) and fully connected classifiers. We use the SVM classifier with the Radial Basis Function (RBF) (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e) kernel. For our fully connected classifier, we designed a simple fully connected network, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, leveraging a pre-trained model for feature extraction. It comprises a fully connected layer with dropout (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e) for regularization, followed by a softmax layer for the final classification. It includes two dense layers and a dropout layer, a widely used regularization technique, forming a typical and basic classifier design.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e1. Nasal Endoscopy Image Dataset\u003c/h2\u003e \u003cp\u003eAfter conducting endoscopic imaging on 150 patients, we selected good-quality images from 46 patients. 18 images were taken from 18 individuals in the normal group, and 74 images were taken from 28 individuals in the allergic rhinitis group, totaling 92 images. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e presents the images obtained for the dataset. Notably, our data set has a significant data imbalance problem. To address this issue, we employed meticulous and thoughtful techniques during image preprocessing and analysis to ensure the reliability and validity of our findings despite the imbalance. The varying imaging distances caused scale differences in nasal cavity images, which the Inception model effectively handled using multiple kernel sizes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2. Comparison Result of Lab Color Distributions between Normal and Allergic Rhinitis Groups\u003c/h2\u003e \u003cp\u003eWe compared the color distributions in the Lab color space between the normal and allergic rhinitis groups using 92 images, as detailed in the Methods section. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e clearly demonstrates the differences between these two distributions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn channel a, the normal group exhibits a higher proportion of histogram values of 16 or higher compared to the allergic rhinitis group. Conversely, in channel b, the allergic rhinitis group shows a higher proportion of histogram values of 0 or lower than the normal group. These observations confirm that the color distribution in the inferior turbinates differs between the normal and allergic rhinitis groups.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e3. Experimental Strategies for Addressing Class Imbalance Problem\u003c/h3\u003e\n\u003cp\u003eAs previously mentioned, the dataset suffers from severe class imbalance. To address this issue, we implemented three strategies in our experiments. Firstly, we repeated experiments by the following cross-validation: randomly dividing the data in ratios of 7:3, 6:4, and 5:5 to create training and validation datasets. The proportion and class weight of data for each split is detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Secondly, we conducted additional experiments employing class weights. For each class weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{w}}_{i}\\)\u003c/span\u003e\u003c/span\u003e, we calculate as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{w}_{i}=\\frac{T}{C\\times\\:{C}_{i}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{T}\\)\u003c/span\u003e\u003c/span\u003e is the total number of images, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{C}\\)\u003c/span\u003e\u003c/span\u003e is the total number of classes, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{C}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the number of images in the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e-th class on training data set. Finally, we adapted F1-score as well as accuracy as performance metrics. F1-score is a statistical measure for a binary classification model when the class distribution is uneven.\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\u003eNumber of images and class weight per class by data split ratio.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7:3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6:4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5:5\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTrain\u003c/p\u003e \u003cp\u003e(weight)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNormal class\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12 (2.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10 (2.70)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9 (2.55)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMast class\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e51 (0.62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e44 (0.61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e37 (0.62)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNormal class\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMast class\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e37\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\u003eExperiments with the SVM classifier are conducted with parameter C set at 1, 10, 100, 1000, and 10000, and gamma at 0.1, 0.01, 0.001, 0.0001, and 0.00001. For the fully connected classifier, we use a learning rate of 0.00001, cross-entropy as the loss function, and Adam optimization as the optimizer.\u003c/p\u003e\n\u003ch3\u003e4. Experimental Results on Classification\u003c/h3\u003e\n\u003cp\u003eFirst, we compared the performance of histogram features and CNN features using the SVM classifier. Tables\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e represent the experimental results for the SVM classifier applied to histogram features and CNN features, respectively. Additionally, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e includes the performance across three different color spaces. For each image, we extracted 100 patches and performed classification individually on each patch. Final classification results per image were obtained through majority voting among the patches, ensuring robust classification by aggregating patch-level predictions.\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\u003ePerformance of histogram features in ab color space using SVM classifier.\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=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData split ratio\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e7:3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003e6:4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e \u003cp\u003e5:5\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass weight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e89.66\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e89.66\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e89.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e89.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e86.96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.9388\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9362\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9375\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.9355\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9351\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9167\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 \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\u003ePerformance of CNN features in RGB, Lab, and ab color spaces using SVM classifier.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eData split ratio\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003e7:3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e6:4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e5:5\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eClass weight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eRGB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e88.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e88.69\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e86.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e86.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e87.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e84.63\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.928\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.9286\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.9188\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9267\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.9233\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLab\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e83.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e81.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e82.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e82.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9077\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9052\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.8987\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.8986\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAb\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e82.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e81.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e81.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e84.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e84.26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.8926\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.8924\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9113\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.9072\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\u003eTables\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e demonstrate the improvement in performance as the proportion of training data increases. Notably, both tables indicate that the application of class weights had no significant effect. Furthermore, Histogram-based features consistently outperformed CNN-extracted features when classified by SVM, likely due to their direct and targeted representation of color distribution differences, which are critical in identifying allergic rhinitis conditions from limited training samples. All the models with class weights showed a slight decrease in performance except for the model using CNN features in the 7:3 split on RGB color space. Regarding color spaces, we can compare the ab color space, which solely contains color components, against Lab and RGB, both of which include color and brightness components. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows that it is difficult to determine the superiority of ab or Lab; however, the RGB color space was clearly superior to both spaces in the case of CNN features.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e was produced by selecting the highest values from those obtained by varying the SVM parameters, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. The highest accuracy of 89.66% and an F1-score of 0.9388 were achieved with C\u0026thinsp;=\u0026thinsp;100 and gamma\u0026thinsp;=\u0026thinsp;0.001.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSecondly, we compared the performance of CNN features across three color spaces using fully connected networks, with data divided in ratios of 7:3, 6:4, and 5:5, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Multiple data splits were employed to ensure robustness and consistency of results under varying training and testing conditions.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of CNN features in RGB, Lab, and ab color spaces using fully connected networks.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eData split ratio\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003e7:3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e6:4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e5:5\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003eClass weight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ew/o\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003ew\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eRGB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e86.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e93.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e84.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e93.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e80.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e86.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9545\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9545\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.9047\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eLab\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e86.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e92.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e80.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e81.57\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9475\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.8915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.8771\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAb\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e89.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e83.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e89.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e80.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e84.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9387\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.9367\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.9295\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.8915\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.9041\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 highest accuracy of 93.1% and F1-score of 0.9545 were significantly better than those achieved with the SVM classifier. Unlike the SVM case, we observed significant performance improvements with class weights in all cases but one case. This outcome demonstrates that class weights in neural networks effectively address data imbalance problems. Similar to previous experiments, the RGB color space clearly exhibited the highest performance among the color spaces.\u003c/p\u003e"},{"header":"Discussions","content":"\u003cp\u003eAllergic rhinitis, a chronic condition that significantly impacts a patient's quality of life, is currently diagnosed using either non-invasive or invasive methods, both of which are time-consuming and costly. This paper presents a novel approach to a deep learning-based allergic rhinitis diagnosis model that utilizes nasal endoscopic images. This method offers a less time-consuming, cost-effective, non-invasive, and qualitative alternative to the existing diagnostic methods.\u003c/p\u003e \u003cp\u003eAllergic rhinitis affects 10\u0026ndash;20% of the population, with 4\u0026ndash;8% annual growth rates. Despite the surge in patent applications since 2015, advancements in diagnostic technologies have been relatively weak (especially in the AR and nasal obstruction diagnostic/decision category, outsourced patent research, and patent attorneys).\u003c/p\u003e \u003cp\u003eHSV color spaces are alternative representations of the RGB (red, green, blue) color model, designed in the 1970s by computer graphics researchers to more closely align with how human vision perceives color-making attributes (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). The colors of each hue are arranged in a radial slice, saturation consists of dimensions resembling various tints of brightly colored paint, and value mentions dimensions resembling the mixture of those paints with varying amounts of black or white paint. Lab color spaces can be closer to human vision than HSV and consistent regardless of monitor changes or printer color variations.\u003c/p\u003e \u003cp\u003eBy analyzing the inferior turbinate color distribution in the Lab color space, we found that the inferior turbinate color distribution of the normal and allergic rhinitis groups differed. Based on this, we extracted CNN features by extracting histogram features and passing them through a pre-trained Inception v3 model on ImageNet weights (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). We conducted comparative experiments using SVM and fully connected classifiers to analyze which feature extraction method is more efficient and which learning technique performs better. SVM is widely regarded as an effective binary classifier. SVM classifiers have the advantage of effectively determining optimal decision boundaries even when sample sizes are limited, thus ensuring reliable generalization from limited training data. Accordingly, we selected SVM as the representative machine learning classifier.\u003c/p\u003e \u003cp\u003eThe experimental results show that the best accuracy of histogram features with the SVM classifier is 0.8966 and F1-score is 0.9388, while the best accuracy of CNN features with the SVM classifier is 0.8821 and F1-score is 0.928, which shows that histogram features perform better. Using a fully connected classifier on the CNN features achieved the best results, with the best accuracy of 0.9310 and F1-score of 0.9545 (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Although the histogram features performed better than those extracted from the CNN trained for general purposes, they require manual effort in the inferior turbinate detection process. Although CNN features perform worse than histogram features, we found that better performance can be achieved by applying a fully connected classifier with class weights (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIt is important to acknowledge the limitations of this study, which include factors such as the nasal endoscopy angle, color adjustment, and analysis program. These limitations highlight the need for continued research and innovation in these areas, underscoring the potential for further advancements in allergic rhinitis diagnosis and treatment.\u003c/p\u003e \u003cp\u003eWe adopted an alternative approach to minimize this bias as much as possible, conducting multiple experiments with different random split ratios (e.g., 7:3, 6:4, 5:5). We ensured patient-wise separation to minimize overlap. Although this is not a traditional n-fold cross-validation method, it represents the most appropriate strategy given the characteristics of our dataset. Furthermore, for this study, we demonstrate that our deep learning-based feature extraction and classifier architecture outperforms conventional hand-crafted features and machine learning-based classifiers.\u003c/p\u003e \u003cp\u003eAlthough this is not a traditional n-fold cross-validation method, it represents the most appropriate strategy given the characteristics of our dataset. Furthermore, this study demonstrated that our deep learning-based feature extraction and classifier architecture outperforms conventional hand-crafted features and machine learning-based classifiers.\u003c/p\u003e \u003cp\u003eIn general, when the amount of data is limited, confidence intervals are appropriate. However, in the case of neural networks, it is common not to report them when the number of test samples is sufficiently large. In our CNN\u0026thinsp;+\u0026thinsp;FC feature extraction and classification approach, we extract 100 additional patches per image (as described below in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Therefore, testing on just 10 images effectively results in classification being performed on 1,000 patches. The final performance is then evaluated by aggregating the classification results of these patches using a voting mechanism. In other words, since the classifier operates at the patch level, the number of test samples is sufficiently large to evaluate the model\u0026rsquo;s performance reliably. In addition, our results rely primarily on accuracy and F1-score due to our initial experimental scope. We chose accuracy as a clear, easily interpretable measure of overall model performance and F1-score as it effectively balances precision and recall, providing valuable insight, particularly for imbalanced datasets such as ours.\u003c/p\u003e \u003cp\u003eNasal endoscopy is likely to vary slightly from doctor to doctor, which can lead to slight angular variations. A certain depth or distance is important because the amount of light can vary. To reduce these errors, a single ENT surgeon performed the study, and white balance and brightness adjustments were performed under the same conditions before the images were acquired. The image was also acquired with a constant depth control to control the amount of light entering the image. Nevertheless, some issues will inevitably lead to differences, which need further study. One solution is the recent development of artificial intelligence in image adjustment, which can be used to adjust the angle slightly. Additional research is needed.\u003c/p\u003e \u003cp\u003eImage analysis, particularly in medical diagnostics, is advancing rapidly. Our study utilized SVM alongside various variables but explored other image analysis methods, including VGG, Inception, Xception, ResNet, DenseNet, and IncRes.\u003c/p\u003e \u003cp\u003eThe images collected in this study were not captured from a fixed distance. Instead, the distance to the target varied depending on the shape of the patient\u0026rsquo;s nasal cavity and the conditions at the time of capture. This led to scale variations of the nasal cavity within the images. Inception, characterized by its use of multiple kernel sizes, is particularly effective in handling such scale variations of target objects.\u003c/p\u003e \u003cp\u003eVarious models, such as VGG, Xception, ResNet, and Inception, can be considered for CNN-based feature extraction in transfer learning. These models differ in usability, the number of parameters, and the characteristics of the extracted features, and their performance may vary depending on the specific image classification task.\u003c/p\u003e \u003cp\u003eIn this study, we used features extracted directly from a pre-trained CNN rather than fine-tuning the CNN with our dataset. We aimed to demonstrate that general-purpose image features learned from ImageNet could outperform even carefully crafted histogram-based features. Therefore, the objective of this study was to obtain general image features.\u003c/p\u003e \u003cp\u003eDespite these innovations, our findings indicated that newer pre-trained models could have yielded better performance and produced consistent results. For future investigations into mucous membrane analysis, such as those conducted in this study, alternative models might be more effective, or there may be a need to develop new pre-trained models specifically tailored for this application.\u003c/p\u003e \u003cp\u003eIn summary, as demonstrated in this study, the optical analysis of nasal endoscopy images could be a valuable adjunct to non-invasive measurement methods for allergic rhinitis. The potential benefits of this new method include improved efficiency, reduced costs, and enhanced patient comfort. However, further studies are necessary to fully evaluate and validate these new methods.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study introduced a novel approach to diagnosing allergic rhinitis using nasal endoscopy images. Our approach analyzed the color distribution of the inferior turbinates within the LAB color space, extracted important features from endoscopy images using both CNN feature extraction and histograms, and performed classification through SVM and fully connected classifiers. Our findings indicated that while histogram features combined with SVM classifiers showed high accuracy and F1 scores, the best results were obtained using CNN features with a fully connected classifier, achieving 90.8% diagnostic accuracy. This suggests that deep learning frameworks can enhance diagnostic accuracy and efficiency when properly tuned and applied to specific medical imaging tasks. However, the study also recognized limitations due to the inherent variability in nasal endoscopy procedures, such as differences in angle and lighting conditions, which can affect image analysis. Future work will address these challenges by refining image capture consistency and exploring the use of advanced image processing technologies.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAR (Allergic rhinitis), SVM (support vector machines), CNN (convolutional neural network), SPT (skin prick tests), MAST (Multiple Allergen Simultaneous Test), IT (inferior turbinate), RBF (Radial Basis Function), RGB (red, green, blue), HSV (hue, saturation, value), LAB (CIELAB), FC (fully connected layers), VGG (Visual Geometry Group\u003c/p\u003e\n"},{"header":"Declarations","content":"\u003cp\u003eThe authors declare no conflict of interest\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003eE-mail: [email protected] (corresponding author)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eConceptualization: Young Joon Jun\u003c/p\u003e\n\u003cp\u003eData curation: MinHye Kang, JaePil Ko, Young Joon Jun\u003c/p\u003e\n\u003cp\u003eFunding acquisition: Young Joon Jun\u003c/p\u003e\n\u003cp\u003eMethodology\u0026mdash;clinical: Young Joon Jun\u003c/p\u003e\n\u003cp\u003eMethodology\u0026mdash;computing: MinHye Kang, JaePil Ko\u003c/p\u003e\n\u003cp\u003eProject administration: \u0026nbsp;JaePil Ko, Young Joon Jun.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eVisualization: MinHye Kang, JaePil Ko, Young Joon Jun\u003c/p\u003e\n\u003cp\u003eWriting\u0026mdash;original draft: MinHye Kang, Young Joon Jun\u003c/p\u003e\n\u003cp\u003eWriting\u0026mdash;review \u0026amp; editing: JaePil Ko, Young Joon Jun\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eFunding\u003c/strong\u003e: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (2019R1I1A3A01063980)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eEthics committee approval and subject consent\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by SoonChunHyang University Gumi Hospital and Uijeongbu Eulji University Hospital Institutional Review Board (IRB). Patients received informed consent.\u003c/p\u003e\n\u003cp\u003eAll research has been performed in accordance with the Declaration of Helsinki.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePawankar R, Bunnag C, Khaltaev N, Bousquet J. Allergic Rhinitis and Its Impact on Asthma in Asia Pacific and the ARIA Update 2008. World Allergy Organ J. 2012;5(Suppl 3):S212-7.\u003c/li\u003e\n\u003cli\u003eLee K-I, Chung Y-J, Mo J-H. The impact of air pollution on allergic rhinitis. Allergy, Asthma \u0026amp; Respiratory Disease. 2021;9(1).\u003c/li\u003e\n\u003cli\u003eYoo KH, Ahn HR, Park JK, Kim JW, Nam GH, Hong SK, et al. Burden of Respiratory Disease in Korea: An Observational Study on Allergic Rhinitis, Asthma, COPD, and Rhinosinusitis. Allergy Asthma Immunol Res. 2016;8(6):527-34.\u003c/li\u003e\n\u003cli\u003eYang SI, Lee IH, Kim M, Ryu G, Kang SY, Kim MA, et al. KAAACI Allergic Rhinitis Guidelines: Part 1. Update in Pharmacotherapy. Allergy Asthma Immunol Res. 2023;15(1):19-31.\u003c/li\u003e\n\u003cli\u003eSzegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the Inception Architecture for Computer Vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)2016. p. 2818-26.\u003c/li\u003e\n\u003cli\u003eKim YH, Kim K-S. Diagnosis and treatment of allergic rhinitis. Journal of the Korean Medical Association. 2010;53(9).\u003c/li\u003e\n\u003cli\u003eBrozek JL, Bousquet J, Agache I, Agarwal A, Bachert C, Bosnic-Anticevich S, et al. Allergic Rhinitis and its Impact on Asthma (ARIA) guidelines-2016 revision. J Allergy Clin Immunol. 2017;140(4):950-8.\u003c/li\u003e\n\u003cli\u003eSeidman MD, Gurgel RK, Lin SY, Schwartz SR, Baroody FM, Bonner JR, et al. Clinical practice guideline: Allergic rhinitis. Otolaryngol Head Neck Surg. 2015;152(1 Suppl):S1-43.\u003c/li\u003e\n\u003cli\u003eJoko H, Hyodo M, Gyo K, Yumoto E. Chromametric assessment of nasal mucosal color and its application in patients with nasal allergy. Am J Rhinol. 2002;16(1):11-6.\u003c/li\u003e\n\u003cli\u003eBae S, Jun YJ. Optical Analysis of Nasal Endoscopic Images From a Patient With Severe Acute Respiratory Syndrome Coronavirus 2. Journal of Rhinology. 2022;29(2):96-100.\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Shea K, Nash R. An Introduction to Convolutional Neural Networks. ArXiv. 2015;abs/1511.08458.\u003c/li\u003e\n\u003cli\u003eScholkopf B, Smola AJ. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond: MIT Press; 2001.\u003c/li\u003e\n\u003cli\u003eMachine Learning and Its Applications2001.\u003c/li\u003e\n\u003cli\u003eSrivastava N, Hinton GE, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014;15:1929-58.\u003c/li\u003e\n\u003cli\u003eIbraheem NA, Hasan MM, Khan RZ, Mishra PK. Understanding color models: a review. ARPN Journal of science and technology. 2012;2(3):265-75.\u003c/li\u003e\n\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":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Allergic Rhinitis Diagnosis, Nasal Endoscope, SVM, CNN, Classification","lastPublishedDoi":"10.21203/rs.3.rs-5221450/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5221450/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eAllergic rhinitis typically has edematous and pale turbinates or erythematous and inflamed turbinates. While traditional approaches include using skin prick tests (SPT) to determine the presence of AR, It is often not related to actual symptoms, and it is an invasive test. We use deep learning to analyze nasal endoscopy images to investigate a quantitative method for diagnosing allergic rhinitis.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eTraditional machine learning-based diagnostic techniques have relied on structured clinical datasets featuring statistical data such as demographic characteristics, symptom severity, and clinical test results. In contrast, we propose a novel approach to use endoscopy image data to analyze the color distribution in the inferior turbinate region of patients with allergic rhinitis using the CIE-Lab color space and extract the adaptive histogram features that are used to explore and find suitable feature extraction methods and deep learning model architectures.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eOur proposed model achieves a promising diagnostic accuracy of 90.80% for images exhibiting AR symptoms. Future research will expand the dataset to include a broader spectrum of symptomatic and asymptomatic images to enhance model robustness and investigate the potential of optical analysis as a non-invasive diagnostic method for AR.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis study introduced a novel approach to diagnosing allergic rhinitis using nasal endoscopy images. Our approach analyzed the color distribution of the inferior turbinates within the LAB color space, extracted important features from endoscopy images using both CNN feature extraction and histograms, and performed classification through SVM and fully connected classifiers.\u003c/p\u003e","manuscriptTitle":"Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-10 04:55:10","doi":"10.21203/rs.3.rs-5221450/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-27T04:26:48+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-22T11:25:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"17313206295391016482143549553903460397","date":"2025-05-12T05:59:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-25T12:03:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"42519429822309851439618149309498451443","date":"2025-04-08T07:33:50+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-08T05:45:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-07T09:34:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-03-25T11:38:11+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"36f06b30-de68-416e-b88e-58dac0e275b3","owner":[],"postedDate":"April 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":46833627,"name":"Biological sciences/Computational biology and bioinformatics/Image processing"},{"id":46833628,"name":"Biological sciences/Computational biology and bioinformatics/Computational models"},{"id":46833629,"name":"Health sciences/Health care/Diagnosis/Physical examination"},{"id":46833630,"name":"Health sciences/Medical research/Experimental models of disease"}],"tags":[],"updatedAt":"2025-07-14T16:10:51+00:00","versionOfRecord":{"articleIdentity":"rs-5221450","link":"https://doi.org/10.1038/s41598-025-10087-x","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-07-08 15:57:53","publishedOnDateReadable":"July 8th, 2025"},"versionCreatedAt":"2025-04-10 04:55:10","video":"","vorDoi":"10.1038/s41598-025-10087-x","vorDoiUrl":"https://doi.org/10.1038/s41598-025-10087-x","workflowStages":[]},"version":"v1","identity":"rs-5221450","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5221450","identity":"rs-5221450","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}