StyleGAN-based synthetic image augmentation for multi-class otoscopy image classification Authors: Seda Camalan, Carl D. Langefeld, Amy Zinnia, Aaron C. Moberly, Metin N. Gurcan Published: 2026 ABSTRACT Accurate diagnosis of eardrum abnormalities is pivotal for effectively managing various ear conditions. Otoscopy, a non-invasive diagnostic procedure, provides detailed views of the ear canal and eardrum to identify pathologies such as otitis media, tympanic membrane perforations, and other abnormalities. Despite advancements, classifying otoscopy images remains challenging due to lighting variations, motion blur, and diverse abnormalities, exacerbated by limited labeled datasets and subjective assessments. A significant barrier to accurate diagnosis is the lack of an objective evaluation method for eardrums, such as machine learning techniques to classify images as normal or abnormal. Moreover, when processing otoscopy videos, traditional approaches require extensive manual expertise to select representative frames for analysis. This study investigates the efficacy of artificial image augmentation using StyleGAN3 to enhance otoscopy image classification accuracy. Augmented datasets comprising artificial and composite otoscopy images were used to train and validate a ResNet-101 model. A total of 816 artificial and 816 composite images (250-Effusion, 340-Normal, 113-Perforation, and 113-Tympanosclerosis) and 80,685 classic augmented images (24651-Effusion, 33660-Normal, 11187-Perforation, and 11187-Tympanosclerosis) were used. Results demonstrate that integrating StyleGAN3-generated synthetic images significantly improves classification performance. Our approach achieved a mean test accuracy of 0.95 ± 0.03 and an F1-score of 0.96 ± 0.02, surpassing those of traditional augmentation methods, while representing an optimistic upper-bound scenario. In contrast, without any augmentation, the accuracy and F1-score are 0.82 ± 0.03 and 0.78 ± 0.02, respectively, and with classic augmentations, they are 0.87 ± 0.02 and 0.85 ± 0.02, respectively. These findings underscore the potential of GAN-based augmentation to bolster eardrum image classification systems and address challenges in diagnostic accuracy.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56954-z. Introduction Accurate diagnosis of eardrum abnormalities is crucial for effective treatment and management of various ear conditions1. Otoscopy, a common diagnostic procedure, involves examining the ear canal and eardrum to detect pathologies such as acute or chronic otitis media, tympanic membrane perforations, and other abnormalities2. This non-invasive examination is vital for identifying conditions such as middle ear infections, excessive earwax accumulation, foreign bodies in the ear canal, Eustachian tube dysfunction, and, in rare cases, tumors within the ear3. The digital otoscope, equipped with specialized cameras and imaging sensors, captures detailed images of the eardrum, providing healthcare professionals with clear and magnified views essential for accurate diagnosis and medical evaluation. However, classifying otoscopy images into multiple categories poses significant challenges. These challenges include lighting variations, motion blur, the diverse appearance of eardrum abnormalities, and the limited number of labeled training datasets available4. Additionally, the diagnostic process is complicated by the subjectivity of visual assessments and the variability in expertise among medical professionals5. Despite advancements in imaging technology, there remains a critical need for automated and reliable classification systems to support clinicians because the otoscopy diagnosis accuracy in primary care settings is between 30% and 67.5%6, and for otologists and pediatricians, it is only between 50% and 73%7. Recent studies continue to report substantial variability in diagnostic performance across deep learning models for otitis media classification, underscoring the challenges posed by limited datasets and inter-class similarity8. Addressing these issues is essential to improving diagnostic accuracy, reducing the risk of misdiagnosis, and ensuring timely and appropriate medical interventions. Deep learning methods, particularly convolutional neural networks (CNNs), have shown immense promise in medical image analysis, achieving human-level performance in tasks such as anomaly detection and classification9–11. CNNs are composed of multiple layers that automatically learn hierarchical features from input images, making them particularly effective for image recognition tasks. In otolaryngology, deep learning has been applied to various diagnostic tasks, including the analysis of gene expression profiles, audiological data, and endoscopic images12–15. Leveraging these methods for otoscopy image classification can significantly enhance the accuracy and efficiency of diagnosing ear conditions. Recent studies have demonstrated the efficacy of deep learning in automatically diagnosing otitis media and other ear diseases by analyzing oto-endoscopic images10. These advancements reduce the dependency on manual feature extraction, streamline the diagnostic process, and support healthcare professionals by providing reliable, automated assessments16,17. More recent work has explored lightweight and interpretable CNN architectures, including knowledge distillation and attention mechanisms, to improve performance while maintaining clinical usability18. One of the critical advantages of deep learning in medical imaging is the ability to leverage transfer learning, in which pre-trained CNNs on large natural image datasets are fine-tuned for specific medical tasks19,20. This approach mitigates the need for extensive labeled medical datasets, which are often difficult to obtain. By adapting pre-trained models, researchers can achieve high classification accuracy for medical images with relatively smaller datasets. Studies employing transfer learning for otoscopy image classification have reported impressive results, often surpassing traditional methods in diagnostic accuracy21,22. CNNs and transfer learning heavily rely on data augmentation, a crucial deep learning technique, especially for medical image classification19. Given the often-limited availability of labeled medical data, augmentation helps artificially expand the training dataset by creating variations of existing images through transformations such as rotation, translation, scaling, and flipping. This process effectively enhances the diversity of the training data, allowing the model to generalize better and reducing the risk of overfitting. For instance, in the context of otoscopy image classification, augmenting the dataset with various transformations can help the model become more robust to the variations seen in real-world clinical settings, such as different angles and lighting conditions during image capture. Video-based approaches leverage all frames from otoscopy videos to reduce reliance on manual frame selection; however, these methods still require robust data augmentation and balanced datasets to generalize effectively23. Utilizing data augmentation is particularly beneficial in transfer learning, where pre-trained CNN models (Deep Convolutional Neural Network -DCNN) on large natural image datasets are fine-tuned for otoscopy image classification and segmentation tasks24–27. The augmented data provides a more extensive and varied set of examples, which helps the fine-tuned model adapt better to the nuances of otoscopy images. This not only improves CNN performance but also ensures the model remains efficient even with smaller datasets, which are common in eardrum applications15. Moreover, advanced data augmentation techniques, such as generative adversarial networks (GANs), can generate highly realistic synthetic images, further enriching the training set. This approach is especially valuable for creating datasets representative of rare medical conditions, thereby enhancing the model’s ability to accurately diagnose them28,29. GANs have been successfully applied to various medical imaging tasks, including the generation of skin lesion images30, histopathology slides31, and COVID-19 chest X-rays32. By creating realistic synthetic images, GANs can enhance the training datasets for deep learning models, improving their performance in diagnosing ear infections and other conditions33. In otoscopy, GANs have promising applications for generating large quantities of diverse otoscopic images33–36, particularly when natural data is limited, such as for rare ear pathologies. Suresh et al.33 proposed a study that utilized the StyleGAN2-ADA architecture to develop a generative adversarial network (GAN) for creating synthetic tympanic membrane (TM) images. Validation by human reviewers, comprising otolaryngology–head and neck surgery (OHNS) and pediatrics trainees, demonstrated a sensitivity of 66%, specificity of 73%, and an AUC of 0.69 in distinguishing synthetic from real images. However, limitations include the restricted generation of only normal TM images and the inherent variability in human assessment, underscoring the need for robust external validation and development of pathological TM images to fully leverage synthetic data in medical applications. Eseoglu et al.34 conducted a study in which a GAN-based system was designed and trained to generate artificial tympanic membrane (TM) images from a dataset of 10,088 images captured with otoscope devices. The training process involved various GAN architectures, including Wasserstein GAN with Gradient Penalty (WGAN-GP)37, Progressive Growing GAN (PGGAN)38, and StyleGAN39. The synthetic images generated were evaluated by otolaryngologists in a Visual Turing test, achieving high realism scores, indicating that the artificial images were perceived as real by specialists. That study highlights the potential use of GAN-generated images in medical training and classification systems. Lu et al. have also emphasized the importance of selecting informative frames from otoscopy videos, showing that automated frame selection can significantly improve downstream diagnostic performance and better reflect real-world clinical workflows40. For this study, we selected a deep convolutional neural network (DCNN) due to its proven ability to automatically extract hierarchical image features and perform multi-class classification with high accuracy, particularly when combined with transfer learning. The DCNN architecture (ResNet-101) leverages residual connections to enable very deep networks without vanishing gradients. To generate realistic synthetic eardrum images, we employed StyleGAN3, a style-based generative adversarial network. StyleGAN3 separates the latent space into multiple style layers, allowing fine-grained control over image features and producing high-fidelity images with minimal artifacts. This combination enables effective augmentation of the training dataset, enhancing the classification model’s robustness. In our study, we created four categories (Effusion, Normal, Perforation, and Tympanosclerosis) of eardrum images to augment the composite images and increase the classification accuracy. Our contributions include: Generating pathological eardrum images such as Effusion, Perforation, and Tympanosclerosis using StyleGAN3.Examining the effect of artificial image augmentation on different stages of the training, validation, and testing phases.Using fewer augmented images to train the model effectively.Comparing the results of classic augmentation techniques with those obtained from artificial image augmentation. Generating pathological eardrum images such as Effusion, Perforation, and Tympanosclerosis using StyleGAN3. Generating pathological eardrum images such as Effusion, Perforation, and Tympanosclerosis using StyleGAN3. Examining the effect of artificial image augmentation on different stages of the training, validation, and testing phases. Examining the effect of artificial image augmentation on different stages of the training, validation, and testing phases. Using fewer augmented images to train the model effectively. Using fewer augmented images to train the model effectively. Comparing the results of classic augmentation techniques with those obtained from artificial image augmentation. Comparing the results of classic augmentation techniques with those obtained from artificial image augmentation. By leveraging advanced machine learning techniques, such as deep learning and data augmentation, we aim to enhance the precision and reliability of eardrum abnormality classification in otoscopy images. This study aims to demonstrate the potential of GAN-based synthetic data generation in enhancing the performance of medical image classification systems, particularly in scenarios where obtaining sufficient original data is challenging. Eardrum classification The methodology section of this study presents a comprehensive experimental setting to address the classification of composite and still otoscopy images obtained from otoscopy video frames. This section explains the critical stages, including data acquisition, data generation, experimental settings, the classification model, and the performance metrics. Dataset All the data were collected by three ENT physicians from three hospitals: The Ohio State University (OSU) and Nationwide Children’s Hospital (NCH) in Columbus, Ohio, and Vanderbilt University Medical Center in Nashville, Tennessee. The database was created with high-resolution digital otoscope videos for both adult and pediatric patients, accompanied by diagnostic labels, with Institutional Review Board (IRB) approval and patient consent. The collected videos do not include identifying protected health information, ensuring patient anonymity. The videos were captured using a high-definition (HD) video otoscope (JEDMED Horus + HD Video Otoscope, St. Louis, MO), an all-in-one handheld digital otoscope with a liquid-crystal display (LCD) viewfinder for user-controlled video capture. This device allows for immediate on-screen review of the image sequences. After each examination, the videos were promptly stored on a secure server. The video frames were originally 1440 by 1080 pixels and recorded in the MPEG-4 video container format. The database covers most major eardrum diagnostic categories, including normal, acute otitis media (AOM), middle ear effusion, perforated eardrum, myringitis, tympanosclerosis, retracted eardrum, cholesteatoma, and eardrum with a ventilation tube in place. Some categories, such as AOM, were excluded due to insufficient sample sizes for effective training of our algorithms. Four categories of otoscopy examination videos with relatively large numbers of examples were included: effusion, eardrum perforation, tympanosclerosis, and normal (healthy). An example of each category’s video frames is shown in Fig. 1. Fig. 1Examples of each category key frame from the otoscopy videos. Examples of each category key frame from the otoscopy videos. As seen in Fig. 1, a normal eardrum is smooth, pearly white or gray, translucent, and intact, with a visible light reflex and clear bony landmarks. A middle ear effusion appears slightly retracted with a yellow or amber coloration of fluid behind it, with or without air-fluid levels or bubbles. An eardrum perforation is a hole or tear in the tympanic membrane, often caused by infection, trauma, or prior placement of a tympanostomy tube. Tympanosclerosis is characterized by white, chalky patches within the eardrum, resulting from the calcification of tissues, often following repeated infections or chronic inflammation. In total, the dataset consisted of 474 cases (otoscopy videos) collected across four diagnostic categories. The class distribution was as follows: Normal (134 cases), Effusion (206 cases), Perforation (66 cases), and Tympanosclerosis (68 cases). Composite images were generated from these videos using the OtoXNet pipeline, and these composite images served as the baseline dataset for all experiments prior to the application of classical and artificial augmentation techniques. The number of composite images exceeds the number of cases because some ears produced multiple composite images, depending on video length and frame variability. Image processing No additional image preprocessing steps, such as denoising, contrast enhancement, histogram equalization, or color normalization, were applied prior to classification. Composite images generated by the OtoXNet pipeline and artificial images generated by StyleGAN3 were used directly as inputs to the classification model. The only preprocessing step was resizing all images to match the classification network’s input dimensions. This approach was chosen to preserve the original visual characteristics of the composite and synthetic images and to avoid bias introduced by preprocessing. Transfer learning with DCNN Transfer Learning (TL)41 is a technique that leverages knowledge gained from solving one problem and applies it to a different task42,43. In the context of DCNNs, this involves using the weights from a model trained on an initial problem to initialize a new model for the target problem by changing the last few layers (generally three layers: dropout, fully connected, and softmax) of the pre-trained model. Although the similarity between source and target tasks and domains affects performance, it is generally more effective than training a model from scratch by freezing initial layers, which mostly extract smaller regions and patterns within the region44. In this study, we utilized the DCNN developed for the ImageNet LSVRC-201545 contest, which classified 1.2 million images into 1000 categories. We employed the ResNet-10146 architecture, which achieved top 1 and top 5 error rates of 19.87% and 4.60%, respectively, in localization and won classification and detection competitions. Also, it performed well in the COCO-201547 with detection and segmentation competitions. The DCNNs were trained in MATLAB 2020a using the Deep Learning Toolbox on a high-performance computer (HPC) equipped with 128 GB of RAM and a 16 GB NVIDIA Tesla P100 PCI-E GPU. ResNet-101 consists of 101 layers, including rectified linear units, max-pooling layers, and residual connections. It was originally trained on ImageNet with an input ROI size of 224 × 224 pixels RGB channel. We applied fine-tuning instead of standard transfer learning, which involves re-training the fully connected layers on new data. We froze the weights of the first 291 layers (out of 347 in MATLAB R2020a) to retain generalized features for transfer learning, while retraining the remaining convolutional and fully connected layers for our specific task. Additional dropout methods were applied in the fully connected layer to reduce overfitting. For our eardrum disease classification task in otoscopy, we added an extra fully connected layer (fc4) to replace the original 1000-node fully connected layer with a 4-node softmax classifier. All the weights and biases in the fully connected layers were randomly initialized for re-training. We also modified the cross-entropy loss function to a weighted cross-entropy loss before fine-tuning. This weighted classification layer computes the weighted cross-entropy loss for classification problems, measuring the error between two continuous random variables. For prediction probabilities y and ground-truth labels t, the weighted cross-entropy loss is defined as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:loss=-\frac{1}{N}\sum\limits_{n=1}^{N}\sum\limits_{c=1}^{C}{w}_{c}{t}_{nc}log\left({y}_{nc}\right)\:$$\end{document} where N is the number of observations, C is the number of classes, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{c}$$\end{document} denotes the class weight for the class c. Generating synthetic eardrum images The main challenge in training the network described in Sect. 2.2 is the limited availability of a large, labeled training dataset. To address this problem and enhance the accuracies in the eardrum classification task, we generated synthetic images in three ways: (1) by creating composite images (Sect. 3.1), (2) by using traditional augmentation techniques (Sect. 3.2) involving various known image manipulations on existing data examples, and (3) by generating new examples learned from the data using generative models (Sect. 3.3). We begin with an overview of stitching video frames, then standard data augmentation techniques, and then generating artificial eardrum images using a StyleGAN3 generative adversarial network. As shown in Fig. 2, we have used otoscopy video frames to generate composite and artificial images using machine learning techniques. Fig. 2Synthetic image generation from the same video pool. Synthetic image generation from the same video pool. Composite images Image stitching techniques in computer vision generate composite images using two main methods: the direct approach and the feature-oriented method. The direct approach calculates camera parameters and transformation matrices by minimizing intensity differences in overlapping regions, but it struggles with scaling, illumination changes, and noise. The feature-oriented method detects and compares features such as edges and points, using descriptors to match key features across images. For the image stitching algorithm, we used Image Composite Editor (ICE) 2.0, a free, non-commercial software package developed by the Microsoft Research Computational Photography Group. ICE is an advanced panoramic image stitcher that combines multiple images with overlapping fields of view to create a high-resolution image or segmented panorama through a multi-step process48. This process includes feature detection, feature matching, transformation estimation, image warping, and blending. Given a set of overlapping photographs of a scene taken from a single camera location, ICE creates a high-resolution panorama that seamlessly combines the original images. For our composite otoscopy images, we use the automatically selected otoscopy video frames from the semantic segmentation method20. Binol et al. addressed the challenge of analyzing video sequences rather than single frames or composite images, focusing on selecting informative frames while eliminating blurry and glare-affected frames. The new OtoXNet approach generates two composite images from frames likely to contain eardrum pathology, transforming the classification problem into a standard image classification task. This method does not require frame-level labels but does use entire otoscope videos, defines a blurriness index, identifies meaningful eardrum regions using α-shapes, and automatically generates composite images. The study highlights the potential for improved diagnosis of eardrum abnormalities using these techniques instead of a single key frame. Classic augmentation CNNs, even with a small architecture, have thousands of trainable parameters. Insufficient data, especially in deep networks with multiple layers, increases the risk of overfitting. Data augmentation is a common strategy to reduce overfitting by artificially expanding the dataset49. Classic augmentation techniques for grayscale images primarily involve affine transformations such as translation, rotation, scaling, flipping, and shearing50. To maintain the eardrum’s characteristics, we avoided transformations that could cause shape deformation, like shearing. Each image was rotated from − 15 to 15 degrees at 3-degree intervals, and each output image was scaled to [0.75, 0.80, 0.85, 0.90, 0.95, 1.0, 1.05, 1.10, 1.20], generating 99 additional images per original sample. Additionally, we applied on-the-fly augmentation by randomly flipping the images horizontally and translating them in the x and y directions within a range of [− 50, 50] pixels. Generating artificial images GANs51 are a generative model framework designed to implicitly learn the data distribution from a given set of samples (e.g., images) and generate new samples drawn from that distribution. A generator and a discriminator are trained simultaneously in a competitive setting of the GAN architecture. The generator aims to produce images indistinguishable from real ones, while the discriminator attempts to differentiate between real and generated images. Among the advancements in GANs, StyleGAN348 stands out for its ability to generate high-resolution images with remarkable fidelity and consistency. Introduced by NVIDIA, StyleGAN3 improves upon its predecessors by addressing issues related to image quality and artifacts through better control of the latent space and enhanced architectural features. This model leverages style-based synthesis, enabling fine-grained control over generated images, particularly useful for tasks that require specific artistic or photorealistic styles. As a result, StyleGAN3 has set new benchmarks in generative modeling, enabling applications ranging from digital art creation to realistic avatar generation. To generate otoscopy images using StyleGAN352, we trained the model with images obtained from video frames. We trained the model separately for each abnormality category and for normal cases. Not all video frames were used because otoscopy video frames present many challenges, such as ear canal views, motion blurriness, and varying lighting conditions. We removed the challenging frames and applied classic augmentation methods to increase the number of training frames. Between forty and fifty thousand frames were used to train StyleGAN3 for each category. The model was trained on a high-performance computer with a 112 CPU, an Nvidia A100 GPU, and 80 GB of RAM. We used Python 3.8, PyTorch 1.9.0, and the CUDA toolkit 11.1 for implementation. All images were resized to a fixed resolution of 512 × 512 pixels prior to training. Each StyleGAN3 model was trained for 5000 kimg using the Adam optimizer with default StyleGAN3 hyperparameters. The batch size was set to 32 due to GPU memory constraints. Training convergence was monitored visually and quantitatively by inspecting generated samples at regular intervals. The final model checkpoint was selected based on visual fidelity and the diversity of generated images, while avoiding obvious artifacts or mode collapse. This class-specific training strategy ensured that artificial images reflected realistic intra-class variability while maintaining diagnostic relevance. Fig. 3Composite, Classic Augmented Composite Images, and Artificially Generated Image examples of the image dataset for each category. Examples are intended to illustrate both representative generated images and the variability of artificial samples; they are not selected solely to show the most visually realistic outputs. Composite, Classic Augmented Composite Images, and Artificially Generated Image examples of the image dataset for each category. Examples are intended to illustrate both representative generated images and the variability of artificial samples; they are not selected solely to show the most visually realistic outputs. Figure 3 provides examples for each category and generative method. Normal and abnormal categories are placed in different columns, while generative methods are shown in different rows. The first row shows examples of composite images; the second row shows examples of classic augmentation applied to composite images, and the third row shows artificial images generated by StyleGAN3. While composite images provide a full view of the eardrum by stitching together several video frames, the artificial images appear realistic, although some exhibit blurriness, and include the ear canal surrounding the eardrum. The StyleGAN3 models were trained prior to the cross-validation splitting using the available dataset for each category. Synthetic images were subsequently generated and incorporated into the classification experiments according to the defined experimental settings. While classification data splits were performed strictly at the case (ear) level to prevent overlap between the training and test sets, we acknowledge that training generative models on the full dataset may introduce a risk of indirect information leakage. This consideration is further discussed in the Discussion section. Experimental settings We designed five experimental settings to investigate the effect of artificial image augmentation in dataset splits. As illustrated in Fig. 4, the first experiment was designed exclusively for composite images. All training, validation, and test data splits contained only composite images, with percentages of 70%, 17.5%, and 12.5%, respectively, which is the default data split setting. We combined composite and artificial images for the second, third, and fourth experimental designs with the same split ratios. Each category’s sample count doubled by adding the same number of artificial images to composite images, except for the first and last experiments. In the second experiment, both composite and artificial images were included in all training, validation, and test sets. In the third experiment, composite and artificial images were included in the training and validation sets, while the test set contained only composite images. In the fourth experiment, composite and artificial images were included in the training set. The validation and test sets consisted exclusively of composite images, with no artificial images included. For the last experiment, only classical augmentation techniques were applied to the composite images in all dataset splits. These experimental settings allow isolation of the impact of artificial image augmentation at different stages of model training and evaluation. For all experiments, models were trained for 10 epochs using the Adam optimizer53, with an initial learning rate of 3 × 10^−4 and a reduction factor of 0.8 every 2 epochs. Early stopping was applied with a threshold of 1 × 10^−5, reserving 10% of the training data for validation. We used a batch size of 64 on a single NVIDIA Tesla P100 PCI-E GPU. An eight-fold cross-validation scheme was used to investigate the effect of augmentation methods across different dataset splits. To ensure independence between training, validation, and test sets, all data splits were performed at the case (ear) level rather than the image level. In other words, all images, including composite and artificial images generated from a given ear, were assigned exclusively to a single fold in the eight-fold cross-validation. This prevents images derived from the same ear from appearing in both training and test sets, thereby avoiding potential overestimation of model performance. Each fold had a balanced distribution of cases across all categories, ensuring comparable class representation during training and evaluation. For each experimental setting, we first split the dataset at the case (ear) level into 70% for training, 17.5% for validation, and 12.5% for testing. Within each training set, a further 10% of cases were reserved as an internal validation set for early stopping. Importantly, these internal validation cases were drawn exclusively from the training set and did not overlap with the main validation or test sets. An eight-fold cross-validation scheme was applied at the case level to the entire dataset, ensuring that each ear appeared in exactly one fold. This approach prevents any overlap between training, validation, and test cases, avoiding data leakage while enabling robust evaluation of model performance across folds. For example, in the third experimental setting, the composite image split for training, validation, and testing is 70%, 17.5%, and 12.5%, respectively. However, since artificial images were used only in the training and validation sets, these sets were split at a 80:20 ratio. Similarly, in the fourth experimental setting, composite images were split according to the default ratio, but all artificial images were included in the training set. To evaluate the performance of the classification model and compare the experimental results, accuracy, F1-score, precision, and recall were used with the following formulas, where TP means true positive, TN means true negative, FP means false positive, and FN means false negative:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:F1-Score=\frac{TP}{TP+0.5\left(FP+FN\right)}\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Precision=\:\frac{TP}{TP+FP}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:Recall=\:\frac{TP}{TP+FN}$$\end{document} Fig. 4Experimental setup to investigate the effect of artificial image augmentation. This figure summarizes the inclusion of composite, artificial, and classically augmented images across training, validation, and test sets for all five experimental settings. Experimental setup to investigate the effect of artificial image augmentation. This figure summarizes the inclusion of composite, artificial, and classically augmented images across training, validation, and test sets for all five experimental settings. Statistical analysis To assess whether differences in model performance across experimental settings were statistically significant, we used a randomized complete block design, treating each cross-validation fold as a block and the experimental setting as the main factor. This approach accounts for variability across folds while enabling paired comparisons between experiments. For each outcome (training, validation, test, and case-level test accuracy), an overall test of differences across experiments was first conducted. When statistically significant, pairwise comparisons between experimental settings were performed using Fisher’s protected least significant difference (i.e., Fisher’s protected LSD) test. A significance level of p<0.05 was used for all analyses. This framework allows for rigorous comparison of augmentation strategies while accounting for the correlated structure of cross-validation results. Results In this study, we investigated the effect of using artificial eardrum images as augmentation data for classifying different eardrum pathologies. As outlined in Sect. 3.4, we evaluated the classification results to understand the impact of various augmentation approaches. Multiple composite images were generated for each video, based on its length and frame variations, using the OtoXNet model. Since a single video recorded from one eardrum could produce multiple images, two separate measurements were defined. The first measurement evaluates classification performance at the individual image level and is referred to as “Test” throughout the paper. The second measurement evaluates performance at the eardrum (case) level and is referred to as “Case Test.” For the Case Test, all images associated with a given ear were aggregated, and the final case-level prediction was determined by majority voting across the image-level predictions. The case was considered correctly classified if the majority predicted class matched the ground-truth diagnosis of the corresponding eardrum. For example, if a single ear (case) produced five composite images and three were classified as ‘Effusion,’ the entire case was labeled as ‘Effusion’ for evaluation purposes. “Image Test” refers to image-level performance, while “Case Test” represents case-level evaluation obtained by aggregating predictions across all images from the same ear using majority voting. Table 1 presents the results for each experimental setup regarding composite image classification sets. The mean and standard deviation of accuracy results for eight-fold cross-validation are listed. Table 1Accuracy results of composite image classification for training, validation, and test for each experimental setting.SetupTrainValidationImage TestCase TestExperiment 10.97 ± 0.010.89 ± 0.030.82 ± 0.030.88 ± 0.03Experiment 20.99 ± 0.000.95 ± 0.010.95 ± 0.030.93 ± 0.03Experiment 30.99 ± 0.000.94 ± 0.020.89 ± 0.030.88 ± 0.03Experiment 40.99 ± 0.000.90 ± 0.010.90 ± 0.030.89 ± 0.02Experiment 51.00 ± 0.001.00 ± 0.000.87 ± 0.020.90 ± 0.04 Table 1Accuracy results of composite image classification for training, validation, and test for each experimental setting.SetupTrainValidationImage TestCase TestExperiment 10.97 ± 0.010.89 ± 0.030.82 ± 0.030.88 ± 0.03Experiment 20.99 ± 0.000.95 ± 0.010.95 ± 0.030.93 ± 0.03Experiment 30.99 ± 0.000.94 ± 0.020.89 ± 0.030.88 ± 0.03Experiment 40.99 ± 0.000.90 ± 0.010.90 ± 0.030.89 ± 0.02Experiment 51.00 ± 0.001.00 ± 0.000.87 ± 0.020.90 ± 0.04 Accuracy results of composite image classification for training, validation, and test for each experimental setting. Overall, the results show that using artificial eardrum images as augmentation data generally improves classification performance. Specifically, Experiment 2 achieved the highest test accuracy of 0.95 with a standard deviation of 0.03, indicating a significant improvement over the “Just Composite” setup, which had a test accuracy of 0.82 ± 0.03. However, this setting includes artificial images in the test set and therefore represents an optimistic, upper-bound scenario rather than a realistic clinical deployment. From a clinical perspective, Experiments 3 and 4 are more representative, as artificial images are restricted to the training phase (and validation phase in Experiment 3), while test data consist exclusively of composite images. Under these conditions, both approaches yield improved performance compared to the baseline, with Experiment 4 achieving a test accuracy of 0.90 ± 0.03 and Experiment 3 achieving 0.89 ± 0.03. The similarity between these results suggests that including artificial images in the validation set does not substantively influence generalization to real test data. Interestingly, the Experiment 5 setup, which used traditional augmentation techniques alongside composite images, achieved perfect accuracy of 1.00 ± 0.00 on both the training and validation sets. However, its test accuracy was 0.87 ± 0.02, indicating possible overfitting to the training data. This is further supported by the “Case Test” results, where Experiment 5 achieved an accuracy of 0.90 ± 0.04, showing slight improvement but not as high as other synthetic-based approaches. Figure 5 illustrates the distribution of fold and case accuracies across different experimental settings for composite image classification. Experiment 2 (Composite + Artificial) demonstrates the highest performance with a median accuracy of 0.97, a lower quartile (Q1) of 0.94, and an upper quartile (Q3) of 0.98. This indicates a high level of consistency and robustness across folds. In contrast, Experiment 1 (Just Composite), with a median accuracy of 0.70, a Q1 value of 0.67, and a Q3 value of 0.75, exhibits the lowest overall performance and greater variability, suggesting that relying solely on composite datasets without augmentation is less effective. Experiments 3 (Composite + Artificial (Train & Validation)) and 4 (Composite + Artificial (Train)) also perform well, with median accuracies of 0.89 and 0.89, respectively. Experiment 3 has a Q1 value of 0.87 and a Q3 value of 0.96, while Experiment 4 shows Q1 and Q3 values of 0.88 and 0.93, respectively. These results highlight the benefits of using artificial images throughout both training and validation stages. Experiment 5 (Composite + Classic Augmentation) presents a moderate performance with a median accuracy of 0.86, a Q1 of 0.85, and a Q3 of 0.89, indicating that while classic augmentation techniques are beneficial, they may not be as effective as artificial image augmentation with GANs. Fig. 5Composite image test accuracy values (a) for folds and (b) for cases with five different experimental sets. Composite image test accuracy values (a) for folds and (b) for cases with five different experimental sets. These findings underscore the importance of advanced augmentation techniques in improving classification accuracy and consistency in medical image analysis. This setup’s wide distribution in the box plot underscores Experiment 1’s inconsistency. Statistical analysis of these results suggests that while both augmentation approaches improve model performance over the Experiment 1 setup, the Experiment 2 and 4 setups tend to achieve higher mean accuracies with slightly greater fold-to-fold variability but greater test-case consistency than the Experiment 5 setup. These results highlight the effectiveness of synthetic augmentation in enhancing the accuracy and robustness of eardrum pathology classification. Table 2Precision, Recall, and F1-score results of composite image classification for each experimental setup.SetupPrecisionRecallF1-scoreExperiment 10.77 ± 0.020.84 ± 0.050.78 ± 0.02Experiment 20.95 ± 0.030.98 ± 0.010.96 ± 0.02Experiment 30.87 ± 0.020.88 ± 0.050.87 ± 0.03Experiment 40.90 ± 0.000.89 ± 0.040.89 ± 0.02Experiment 50.84 ± 0.020.88 ± 0.030.85 ± 0.02 Table 2Precision, Recall, and F1-score results of composite image classification for each experimental setup.SetupPrecisionRecallF1-scoreExperiment 10.77 ± 0.020.84 ± 0.050.78 ± 0.02Experiment 20.95 ± 0.030.98 ± 0.010.96 ± 0.02Experiment 30.87 ± 0.020.88 ± 0.050.87 ± 0.03Experiment 40.90 ± 0.000.89 ± 0.040.89 ± 0.02Experiment 50.84 ± 0.020.88 ± 0.030.85 ± 0.02 Precision, Recall, and F1-score results of composite image classification for each experimental setup. In addition to accuracy measurements, precision, recall, and F1-score are used as evaluation metrics. Table 2 shows the average precision, recall, and F1-score values for each experimental setup. These results are similar to the accuracy results. The higher F1-score is observed on Experiment 2, and Experiment 4 follows it with 0.96 ± 0.02 and 0.89 ± 0.02, respectively. While any kind of augmentation increases success, artificial image augmentation is slightly better than classic augmentation methods. The precision, recall, and F1-score values are illustrated in Tables 3 and 4 by category. For each evaluation metric, Experiment 1 shows that the ‘Normal’ category has higher values than the other experiments. However, after any kind of augmentation, that gap closes. Especially for Experiment 2 and 3 setups, all the categories’ successes are similar and higher, while Experiment 5, which is just classic augmentation, has a difference between the ‘Normal’ and ‘Effusion’ categories. Table 3Each eardrum category and its precision, recall, and F1-score values for each experimental setup, and the averages of these values.PrecisionRecallExp1Exp2Exp3Exp4Exp5Exp1Exp2Exp3Exp4Exp5Effusion0.770.960.890.870.860.830.980.940.900.91Normal0.790.990.890.900.810.860.990.880.900.88Perforation0.770.940.900.870.840.780.990.860.800.85Tympanosclerosis0.740.920.900.860.840.900.960.890.920.88Mean0.770.950.900.870.840.840.980.890.880.88Standard deviation0.020.030.000.020.020.050.010.040.050.03 Table 3Each eardrum category and its precision, recall, and F1-score values for each experimental setup, and the averages of these values.PrecisionRecallExp1Exp2Exp3Exp4Exp5Exp1Exp2Exp3Exp4Exp5Effusion0.770.960.890.870.860.830.980.940.900.91Normal0.790.990.890.900.810.860.990.880.900.88Perforation0.770.940.900.870.840.780.990.860.800.85Tympanosclerosis0.740.920.900.860.840.900.960.890.920.88Mean0.770.950.900.870.840.840.980.890.880.88Standard deviation0.020.030.000.020.020.050.010.040.050.03 Each eardrum category and its precision, recall, and F1-score values for each experimental setup, and the averages of these values. Table 4Each eardrum category and its F1-score values for each experimental setup, and the average precision, recall, and F1-score.F1-scoreAverageExp1Exp2Exp3Exp4Exp5PrecisionRecallF1-scoreEffusion0.790.970.910.880.880.870.910.89Normal0.810.990.880.890.830.870.900.88Perforation0.760.960.870.830.840.860.850.85Tympanosclerosis0.770.940.900.880.850.850.910.87Mean0.780.960.890.870.850.870.890.87Standard deviation0.020.020.020.030.020.010.030.02 Table 4Each eardrum category and its F1-score values for each experimental setup, and the average precision, recall, and F1-score.F1-scoreAverageExp1Exp2Exp3Exp4Exp5PrecisionRecallF1-scoreEffusion0.790.970.910.880.880.870.910.89Normal0.810.990.880.890.830.870.900.88Perforation0.760.960.870.830.840.860.850.85Tympanosclerosis0.770.940.900.880.850.850.910.87Mean0.780.960.890.870.850.870.890.87Standard deviation0.020.020.020.030.020.010.030.02 Each eardrum category and its F1-score values for each experimental setup, and the average precision, recall, and F1-score. Statistical comparison across experimental settings To further evaluate differences between experimental settings, statistical analysis was performed using a randomized complete block design with cross-validation folds treated as blocks. Significant differences were observed across all outcomes, including training, validation, and test performance, as well as case-level test performance (all p<0.05). For training performance, Experiments 1 and 5 were significantly different from all other experiments, while Experiments 2, 3, and 4 were not significantly different from each other. In terms of validation performance, Experiment 5 differed significantly from all other experiments, while Experiments 2 and 3 showed no significant difference between them. For test performance, Experiments 1, 2, and 5 were each significantly different from the others. Notably, Experiments 3 and 4 did not differ significantly, supporting their similar performance under clinically realistic conditions. For case-level test performance, fewer differences were observed. Experiment 2 showed significant differences compared to several other experiments, while Experiment 5 did not differ significantly from the remaining settings. Overall, these results confirm that artificial image augmentation significantly improves performance compared to the baseline (Experiment 1), while differences between clinically realistic augmentation strategies (Experiments 3 and 4) are not statistically significant. Detailed pairwise comparison results for all outcomes are provided in the Supplementary Material (Supplementary Tables S1–S4). Discussion In this study, we explored the effectiveness of artificial image augmentation, specifically using StyleGAN3, in enhancing the classification accuracy of multi-class otoscopy images. The motivation for this research stemmed from the inherent challenges in diagnosing eardrum abnormalities, including lighting variations, motion blur, and limited labeled datasets. By generating synthetic images and augmenting the training dataset, we aimed to overcome these obstacles and build a more robust, reliable classification system. Even though our results showed significant improvements in classification performance, underscoring the potential of GAN-based augmentation in medical imaging, there are some challenges and constraints. Challenging artificial images One of the key challenges in utilizing artificial images, particularly those generated by StyleGAN3, is ensuring the realism and clinical relevance. Although StyleGAN3 has demonstrated a remarkable ability to generate high-fidelity images, the generated images can still contain artifacts or unrealistic features that may not be clinically meaningful. This is particularly critical in medical imaging, where subtle details can significantly impact diagnosis and treatment decisions. Therefore, it is essential to continuously evaluate and refine GAN models to improve the quality and realism of synthetic images for clinical training. To address this limitation and further improve the clinical applicability of synthetic data, several strategies can be considered. First, optimizing GAN training procedures, such as increasing training duration, utilizing higher-resolution image synthesis, and ensuring better class balance, may help reduce artifacts and improve anatomical consistency. Second, the use of more advanced architectures or improved regularization techniques could enhance visual fidelity and reduce the presence of unrealistic features. Future work will explore physics-based rendering and the incorporation of anatomy-aware loss functions during StyleGAN training. Third, incorporating expert-driven evaluation methods, such as structured clinician review or Visual Turing tests, would provide a more reliable assessment of clinical realism. Additionally, implementing automated or manual filtering of low-quality synthetic images prior to training may further improve model robustness. These directions are essential to ensure that synthetic image augmentation translates effectively into real-world clinical applications. However, our hypothesis and results show that, although the artificial images are clinically unrealistic, as illustrated in Fig. 6, they are helpful for image augmentation and model training. This observation suggests that the effectiveness of GAN-based augmentation in this context is not solely dependent on photorealism, but rather on its ability to introduce variability in low- and mid-level features such as color distributions, textures, and structural patterns. These variations may act as a form of regularization, improving the model’s robustness and generalization, even when some generated samples lack full anatomical realism. Figure 6 illustrates some of the challenges of the artificial images for each category as follows: Normal: The color of the eardrum is significant for identifying a normal eardrum; however, the shape and pattern are completely meaningless and do not reflect any realistic eardrum appearance.Effusion: Even though the amber color is a very specific feature of effusion and is obvious behind the eardrum, it appears anatomically problematic because it seems to have two malleus bones.Perforation: This eardrum is perforated. However, it seems necessary to clarify that it is indeed an eardrum, as it appears to be a combination of several eardrums.Tympanosclerosis: The color and patterns of tympanosclerosis are easily distinguishable in this image. However, neither the structure nor the shape of the eardrum resembles a real eardrum. Additionally, the non-black area of the image is not even circular or oval, as would be expected if it were obtained from an otoscope. Normal: The color of the eardrum is significant for identifying a normal eardrum; however, the shape and pattern are completely meaningless and do not reflect any realistic eardrum appearance. Normal: The color of the eardrum is significant for identifying a normal eardrum; however, the shape and pattern are completely meaningless and do not reflect any realistic eardrum appearance. Effusion: Even though the amber color is a very specific feature of effusion and is obvious behind the eardrum, it appears anatomically problematic because it seems to have two malleus bones. Effusion: Even though the amber color is a very specific feature of effusion and is obvious behind the eardrum, it appears anatomically problematic because it seems to have two malleus bones. Perforation: This eardrum is perforated. However, it seems necessary to clarify that it is indeed an eardrum, as it appears to be a combination of several eardrums. Perforation: This eardrum is perforated. However, it seems necessary to clarify that it is indeed an eardrum, as it appears to be a combination of several eardrums. Tympanosclerosis: The color and patterns of tympanosclerosis are easily distinguishable in this image. However, neither the structure nor the shape of the eardrum resembles a real eardrum. Additionally, the non-black area of the image is not even circular or oval, as would be expected if it were obtained from an otoscope. Tympanosclerosis: The color and patterns of tympanosclerosis are easily distinguishable in this image. However, neither the structure nor the shape of the eardrum resembles a real eardrum. Additionally, the non-black area of the image is not even circular or oval, as would be expected if it were obtained from an otoscope. The reasons for the unrealistic eardrum images include the complexity of eardrum anatomy, the presence of multiple overlapping structures, and the variability in lighting and angle in the training data. These factors contribute to the generation of synthetic images that may not accurately replicate the intricate details and variations of real eardrums. Importantly, no manual curation or filtering of artificial images was performed prior to training, ensuring that performance gains reflect robustness to imperfect synthetic samples rather than selective inclusion. Fig. 6Clinically unrealistic artificial otoscopy images. Clinically unrealistic artificial otoscopy images. Lack of a turing test for ENT specialists Another limitation of our study is that we did not conduct a Visual Turing test with ENT specialists to assess the realism of artificial images. A Visual Turing test, in which experts are asked to distinguish between real and artificial images, would provide valuable insights into perceptual realism and clinical plausibility. Such expert-based evaluation would be particularly valuable for assessing whether synthetic images preserve clinically relevant diagnostic cues beyond visual similarity. Several prior studies have incorporated such evaluations and reported moderate expert agreement in differentiating synthetic from real otoscopy images Suresh et al.33 and Eseoglu et al.34. In this work, the primary objective of generating artificial images was not to achieve perceptual indistinguishability from real images, but rather to evaluate their effectiveness as a data augmentation strategy for improving classification performance. As a result, we focused on quantitative assessment and downstream task performance rather than expert-based visual evaluation. To provide a quantitative indication of image similarity within our dataset, we computed the Fréchet Inception Distance (FID) between real composite images and generated artificial images for each category. It is important to note that FID scores are not directly comparable across studies because they depend on dataset characteristics, image resolution, preprocessing pipelines, and feature extractors. Therefore, in this study, FID was used solely for relative, within-dataset comparison across diagnostic categories rather than for benchmarking against previously published values. As shown in Table 5, the lower FID scores observed for the Normal and Effusion categories are consistent with their more structured and visually homogeneous appearance, whereas higher FID values for Perforation and Tympanosclerosis likely reflect their greater anatomical variability. Table 5FID score of each category of the real versus generated artificial images.Real imageArtificial imageFID scoreEffusionEffusion generated160.8NormalNormal generated137.8PerforationPerforation generated180.9TympanosclerosisTympanosclerosis generated179.8 Table 5FID score of each category of the real versus generated artificial images.Real imageArtificial imageFID scoreEffusionEffusion generated160.8NormalNormal generated137.8PerforationPerforation generated180.9TympanosclerosisTympanosclerosis generated179.8 FID score of each category of the real versus generated artificial images. Although a Visual Turing test would provide complementary qualitative insight, its absence does not diminish the primary findings of this study, as the artificial images consistently improved classification performance when used as augmentation, even when evaluation was restricted to real composite images. Incorporating expert-based perceptual evaluation remains an important direction for future work, particularly assessing clinical realism and acceptance of synthetic images beyond their utility for model training. Number of artificial and classic augmented images Our study involved generating a large number of artificial images using StyleGAN3, along with traditional augmentation techniques. The use of both types of augmentation aimed to address the limitations of small datasets and to improve the robustness of the classification model. However, the balance between artificial and classical augmented images is crucial. While artificial images can introduce new variations and enrich the dataset, they can also introduce biases if not properly validated. Classic augmentation techniques, on the other hand, ensure that the augmented images remain within realistic and clinically relevant transformations. From a clinical applicability perspective, this distinction reflects an important trade-off. Classical augmentation methods are generally intended to preserve anatomical structure; however, certain transformations, such as excessive geometric distortion or unrealistic intensity variations, may still introduce patterns that are not clinically plausible. In contrast, GAN-based augmentation can generate novel structural variations that may enhance model generalization but do not always correspond to physiologically valid anatomy. Therefore, while GAN-generated images can enrich the diversity of the training data, careful consideration is required to balance performance gains with clinical interpretability and realism. From a clinical deployment perspective, evaluation scenarios in which artificial images are restricted to the training phase, such as Experiments 3 and 4, are more representative, as test data in real practice consist exclusively of real composite images. This setup also helps ensure that any potential artifacts or unrealistic patterns introduced during augmentation, whether from GAN-based or classical methods, do not directly influence model evaluation. One of the key findings of our study is that even with the same number of artificial images, accuracy can be higher than with 100 times more classic augmented images. Artificial images are more effective than classical augmentation not only because they enhance dataset diversity but also because they use memory more efficiently and incur lower computational cost during training54. Generating and storing artificial images requires significantly less memory and storage space than maintaining extensive sets of classically augmented images, making them a more practical solution for large-scale data augmentation. Although classical augmentation remains competitive, as demonstrated in Experiment 5, the slight performance gap observed in clinically realistic settings (Experiments 3 and 4) highlights the complementary role of artificial image augmentation rather than a direct replacement. In the biomedical field, certain augmentation techniques may lack practical relevance; for example, flipping a slice or volume horizontally or vertically is not typically encountered in real clinical settings. Therefore, developing a deeper understanding of data augmentation strategies in biomedicine is crucial to ensure they enhance rather than hinder productivity55. A particularly informative comparison arises between Experiment 3, which incorporates artificial images during training and validation, and Experiment 5, which relies solely on classical augmentation techniques. While the observed performance difference between these two settings is modest, the qualitative implication is significant. Classical augmentation is limited to affine transformations of existing pathologies, effectively ‘remixing’ known data. From a clinical applicability perspective, this distinction reflects an important trade-off. Classical augmentation maintains strict fidelity to real anatomical structures, ensuring that all variations remain clinically plausible. In contrast, GAN-based augmentation can introduce novel structural patterns that may enhance model generalization but do not always correspond to physiologically valid anatomy. Therefore, while GAN-generated images can enrich the diversity of the training data, careful consideration is required to balance performance gains with clinical interpretability and realism. In contrast, artificial augmentation generates novel structural variations, such as different perforation shapes or tympanosclerosis patterns, which do not exist in the training set. This suggests that GANs contribute semantic variability, whereas classical methods contribute only geometric variability. The slightly higher and more consistent performance observed in Experiment 3 indicates that artificial images may contribute additional structural variability beyond what is achievable through label-preserving classical transformations alone. Importantly, this minimal performance gap indicates that the value of artificial image augmentation should not be interpreted solely through accuracy improvements. Instead, its benefit lies in providing a complementary augmentation strategy that can reduce overfitting, balance class-level performance, and improve robustness without requiring large numbers of classically augmented samples. In this context, artificial image augmentation may be particularly advantageous when the dataset size is limited or when certain classical transformations lack clinical relevance. Future studies should explore the optimal ratio of artificial to classical augmented images to maximize classification performance while maintaining clinical validity. By carefully balancing augmentation strategies and effectively managing computational resources, the performance and real-world applicability of deep learning models for otoscopy image classification can be further improved. To further strengthen the robustness of our findings, we performed statistical significance testing across experimental settings using a randomized complete block design. The results confirmed that augmentation strategies significantly improve performance compared to the baseline (Experiment 1). Importantly, no statistically significant differences were observed between Experiments 3 and 4 in test performance, suggesting that restricting artificial images to the training phase provides generalization comparable to including them in both training and validation. This finding reinforces the clinical relevance of these configurations, as they more closely reflect real-world deployment scenarios. While Experiment 2 achieved the highest accuracy, including artificial images in the test set represents an optimistic scenario. The statistical analysis supports the conclusion that the improvements observed in Experiments 3 and 4 are both meaningful and robust, strengthening the validity of artificial augmentation as a practical strategy for clinical applications. Novel application of artificial image augmentation Our study is among the first to explicitly investigate perforation and tympanosclerosis as separate categories in the context of artificial image augmentation for otoscopy image classification. While prior studies have included these pathologies as distinct diagnostic classes, most notably in smartphone-based deep learning systems for otoscopic analysis, their treatment has typically focused on classification performance using real images alone56. In contrast, our work examines the effect of incorporating artificial images into these clinically important but underrepresented categories, aiming to improve data balance and model robustness. By evaluating perforation and tympanosclerosis within an augmentation-driven experimental framework, we aim to better understand how synthetic data can support fine-grained classification of specific eardrum pathologies. This distinction is particularly relevant in otoscopy datasets, where certain pathological conditions occur less frequently and are therefore more susceptible to class imbalance. Our findings suggest that artificial image augmentation can contribute to improved generalization for these categories, highlighting its potential utility in enhancing clinically meaningful classification granularity in otolaryngology. Despite the promising results presented in this study, several methodological limitations should be acknowledged. Because the generative models were trained before cross-validation splitting, indirect information leakage cannot be completely ruled out, and this may have influenced the apparent benefit of synthetic augmentation. In addition, the absence of external validation limits the generalizability of the proposed approach. All experiments were conducted on a single dataset, and model performance may vary across different clinical settings, imaging devices, and patient populations. We also did not include expert perceptual assessment of synthetic image realism, such as a Visual Turing test with ENT specialists. Finally, the analysis was limited to four diagnostic categories and excluded acute otitis media (AOM), a clinically significant and commonly encountered diagnosis. Future work should examine broader pathology coverage, including AOM and other underrepresented or diagnostically challenging tympanic membrane abnormalities, using both real and synthetic data in multi-center validation studies to improve model generalizability and clinical utility. Despite these limitations, the use of synthetic images may offer a useful conceptual parallel to clinical education. Medical trainees are often first exposed to idealized anatomical illustrations before encountering the full variability of real clinical cases. Similarly, synthetic images can provide structured and diverse representations that help models learn fundamental patterns before generalizing to real-world otoscopy data. In this sense, synthetic augmentation should not be viewed as a substitute for real clinical data, but rather as a complementary strategy that may support model training when real examples are limited or imbalanced. Conclusion In conclusion, our study demonstrates the effectiveness of using StyleGAN3-generated artificial eardrum images for augmenting the classification of otoscopy images. By creating artificial pathological images such as Effusion, Perforation, and Tympanosclerosis, we expanded the diversity of our dataset, thereby enhancing the classification accuracy of our deep learning model. The highest performance was observed when artificial images were included in the test set, yielding a mean test accuracy of 0.95 ± 0.03 and an F1-score of 0.96 ± 0.02; however, this represents an optimistic upper-bound scenario. More clinically realistic experiments, in which evaluation was restricted to real composite images, also showed improvement over the baseline, with test accuracies of 0.89–0.90 compared with 0.82 without augmentation and 0.87 with classical augmentation. Our findings highlight the potential of GAN-based synthetic data augmentation as a powerful tool in medical image analysis, particularly in scenarios where original data acquisition is limited or challenging. Comparing our results with existing literature underscores the significant improvement achieved through artificial data augmentation. Traditional methods often struggle with limited datasets and require extensive manual effort for accurate diagnosis, whereas our approach leverages advanced machine learning to streamline and enhance classification. This aligns with previous studies showing the efficacy of deep learning and transfer learning in medical imaging tasks, emphasizing the robustness and scalability of our method. While our study demonstrates promising results, it is essential to acknowledge its strengths and limitations. Using StyleGAN3 to generate synthetic images effectively improved classification accuracy with a relatively small dataset. However, the realism and diversity of GAN-generated images remain areas of ongoing development. Furthermore, external validation and the incorporation of a broader range of pathological conditions are necessary to fully harness the potential of synthetic data in clinical applications. Our research underscores the transformative potential of artificial data augmentation in medical imaging, paving the way for future advancements in automated diagnostic systems for ear conditions. Continued research and development in this area promise to further refine and optimize deep learning and GANs, ultimately enhancing healthcare outcomes and patient care in otolaryngology and beyond. While the highest performance was achieved in experimental settings that included artificial images in the test set, such scenarios may not fully reflect real-world clinical deployment. In practice, models are evaluated exclusively on real patient data. Therefore, results from Experiments 3 and 4, where artificial images were restricted to the training phase, provide a more realistic estimate of model performance. These findings indicate that appropriately applied artificial image augmentation can improve model generalization, without relying on synthetic data during evaluation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1