Automatic Identification of Adenoid Hypertrophy via Ensemble Deep Learning Models Employing X-ray Adenoid Images

Abstract

Adenoid hypertrophy, characterized by the abnormal enlargement of adenoid tissue, is a condition that can cause significant breathing and sleep disturbances, particularly in children. Accurate diagnosis of adenoid hypertrophy is critical, yet traditional methods, such as imaging and manual interpretation, are prone to errors. This study uses an ensemble deep learning-based approach for adenoid classification. It utilizes a unique dataset sourced from Batman Training and Research Hospital. The dataset is composed of masked and non-masked X-ray images. It is used to train and compare the performance of multiple convolutional neural network (CNN) models. By comparing classification accuracy between masked and non-masked datasets, the study reveals the importance of image preprocessing. Six deep learning models—EfficientNet, MobileNet, ResNet50, ResNet152, VGG16, and Xception—are tested, with ResNet50 achieving the highest accuracy (100% on masked images), while Xception performs the worst (65% F1-score). The results indicate that masking significantly enhances the accuracy and reliability of adenoid classification. ResNet50 and EfficientNet show strong generalization capabilities. Conversely, the lower performance of models like Xception highlights the variability in model suitability for this task. This research provides valuable insights into optimizing deep learning models for medical image classification and it advances the field of AI-based adenoid detection.

Introduction

Adenoid hypertrophy is a condition where the adenoid glands, located at the back of the nose, become enlarged. This enlargement can lead to various health issues, particularly by restricting airflow through the nasal passages. The adenoids, part of the immune system, normally shrink after childhood; however, in some cases, they can remain enlarged in both children and adults. Detecting adenoid hypertrophy is important because it can lead to complications like breathing difficulties, sleep apnea, and nasal obstruction. Diagnosing the condition accurately is crucial to prevent these complications and ensure proper treatment [1, 2].

Traditionally, medical professionals rely on methods like endoscopy and X-ray imaging to assess the size and impact of the adenoids. The goal is to determine how much of the nasopharynx (the upper part of the throat behind the nose) the enlarged adenoids are affecting. This is important to assess any blockage that may restrict breathing. However, the process of analyzing images to detect adenoid hypertrophy can be challenging due to variations in imaging techniques, patient positioning, and the radiologist’s experience [3, 4]. Moreover, manual image interpretation can be time-consuming and prone to human error, especially in difficult cases [5].

To address these challenges, automated detection methods are being developed to assist in identifying adenoid hypertrophy more accurately. By using computer-based algorithms, these systems can process medical images and identify the adenoid region without needing extensive manual intervention [6, 7]. Automated systems not only reduce the workload of radiologists but also improve the consistency of diagnoses. These systems can apply image processing techniques that enhance visibility, and it makes it easier to identify the adenoid and determine whether it is enlarged. This advancement can lead to faster and more reliable diagnostic outcomes [8].

Deep learning technologies [9–13] which involve training computers to recognize patterns in data are being utilized to improve the accuracy of detecting adenoid hypertrophy. Deep learning models can be trained to analyze medical images, and they segment the adenoid gland from surrounding tissues. This process, known as segmentation, helps in clearly distinguishing the adenoid from its background. Once the adenoid is segmented, its size and impact on the airway can be assessed more easily. The use of deep learning in this field has shown great potential for improving both the speed and accuracy of diagnosis [14].

The development of automated detection systems using deep learning is a promising advancement in the diagnosis of adenoid hypertrophy. These systems provide a more efficient and reliable way to analyze medical images, offering significant benefits to both patients and medical professionals. As technology continues to evolve, the integration of automated systems in clinical settings could become standard practice. It improves the quality of care and reduces the chances of misdiagnosis. By making the detection of adenoid hypertrophy faster and more accurate, these systems have the potential to enhance patient outcomes and streamline the diagnostic process [15].

The main target of this research lies in the development and application of enhanced deep learning models for the classification and diagnosis of adenoid hypertrophy. The proposed approach integrates advanced CNN architectures with medical imaging data. This integration aims to enhance the accuracy and reliability of adenoid hypertrophy detection. By improving diagnostic precision, it also seeks to reduce diagnostic times. Ultimately, this approach supports early intervention strategies for better patient outcomes. This work not only contributes to the growing body of literature on AI applications in healthcare but also addresses a critical gap in the efficient diagnosis of a condition that significantly impacts the quality of life in both children and adults.

In this study, several key contributions are made that significantly advance the research in adenoid classification using deep learning models:

1. This research leverages a unique adenoid dataset sourced from the Batman Training and Research Hospital. It is marked the first time this dataset has been utilized for scientific analysis. By introducing a novel and previously untapped dataset, the study expands the scope of medical image analysis for adenoid hypertrophy detection. This dataset is valuable because it contains diverse cases of adenoid hypertrophy, representing real-world clinical scenarios. The use of this dataset enhances the generalizability of the model and provides a solid foundation for future studies in adenoid classification and related conditions.

2. To achieve high accuracy in classifying adenoid hypertrophy, this paper employs an ensemble model, combining the predictive strengths of multiple deep learning architectures. Ensemble learning methods have been shown to significantly improve the robustness and precision of models by aggregating the results from various classifiers to minimize errors.

3. A novel aspect of this study is the comparative analysis of adenoid classification using both masked and non-masked datasets. In this paper, the performance of the model on both masked and non-masked datasets will be evaluated to classify images as adenoid or non-adenoid. This comparative analysis will provide insights into the necessity of segmentation in the classification task, and the results will help determine whether segmentation significantly enhances accuracy or if comparable results can be achieved without it.

Related Works

Adenoid hypertrophy (AH) is a common condition in pediatric populations that can lead to obstructive sleep apnea (OSA), mouth breathing, and subsequent craniofacial abnormalities. Traditionally, the diagnosis of AH has relied on manual techniques such as lateral cephalometry and nasal endoscopy, which are often time-consuming, subjective, and reliant on the clinician’s expertise. Recent advancements in artificial intelligence (AI) and deep learning have introduced automated methods to streamline this diagnostic process, improving both the accuracy and efficiency of AH assessment [2].

Zhao et al. [16] proposed a deep learning model based on convolutional neural networks (CNNs) for the automatic assessment of AH using lateral cephalograms. Their system utilized Fujioka’s method, which calculates the adenoid nasopharyngeal (AN) ratio to detect adenoid hypertrophy. By automating the identification of key landmarks, such as the most convex point on the adenoid shadow and the posterior superior edge of the hard palate, their AI model achieved a high diagnostic performance with an accuracy of 91.9%, a sensitivity of 90.6%, and a specificity of 93.8%. This model demonstrated that CNNs can effectively reduce the manual workload associated with cephalometric measurements, while maintaining high diagnostic accuracy.

Rao et al. [17] extended the capabilities of lateral cephalogram-based AH detection by developing AdeNet, a multi-scale local attention deep learning model. AdeNet introduces a local attention mechanism that focuses on spatial and channel-specific information related to the adenoid region, thus enhancing the model’s ability to differentiate between normal and hypertrophic adenoids. By converting the AN ratio calculation task into a keypoint detection problem, the model automates the marking of relevant cephalometric points. This method outperformed traditional CNN-based approaches by achieving a mean squared error (MSE) of 0.0023 and a mean radial error (MRE) of 1.91 on three independent datasets and it demonstrates superior localization of anatomical landmarks.

In another study, Dong et al. [18] explored the application of deep learning for AH diagnosis using cone-beam computed tomography (CBCT). They proposed a hierarchical mask self-attention U-Net (HMSAU-Net) for upper airway segmentation and a 3D-ResNet model for classifying AH. This method capitalized on the three-dimensional structure of the upper airway, and it provides a more comprehensive view of adenoid hypertrophy compared to traditional two-dimensional cephalograms. The system achieved a diagnostic accuracy of 91.2%, with a sensitivity of 97.6% and a specificity of 86.7% that illustrates the potential of CBCT in enhancing the accuracy of AH diagnosis by considering the spatial complexity of the upper airway.

Moreover, Fu et al. [19] developed a lightweight deep learning model, AFFMNet, aimed at classifying adenoid cystic carcinoma (ACC), a malignant tumor with histological similarities to benign conditions like basal cell adenoma (BCA). Although not directly related to AH, this model introduced important innovations in AI-based medical image classification by employing adaptive feature fusion and attention mechanisms. AFFMNet demonstrated exceptional accuracy in patch-level classification (97.37%) and whole slide image classification (92.91%). The architecture and methods used in AFFMNet are potentially applicable to other diagnostic challenges, including AH, where large datasets and high-resolution images could benefit from efficient feature extraction and classification.

In another study by Alshbishiri et al. [3], the authors proposed a deep learning-based method using U-Net for segmenting adenoid glands in X-ray images, and it demonstrates the utility of U-Net in handling variability in size, shape, and position of the adenoid gland. The study achieved a Dice coefficient score of 0.74, which shows the potential for automated methods to assist radiologists by minimizing the subjectivity and variability associated with manual analysis.

Another study by He et al. [2] presented a novel approach using MRI images for AH detection through a deep learning-based model called ADNet. This model aimed to improve on conventional methods, which were previously reliant on X-rays or landmark detection techniques. The authors employed an encoder-decoder architecture, effectively combining local and global feature extraction for accurate assessment. ADNet was shown to achieve high accuracy, with an average radial error of 4.85 pixels and an average successful detection rate of 96% within 15 pixels, indicating its effectiveness in assisting clinicians in analyzing AH.

Collectively, these studies highlight the significant progress made in automating AH diagnosis through AI and deep learning techniques. By employing CNNs, attention mechanisms, and 3D imaging technologies, researchers have developed models that not only reduce the diagnostic burden on clinicians but also enhance the accuracy and reliability of AH detection. These automated approaches are paving the way for more standardized, objective, and scalable diagnostic solutions in both dental and medical practices.

The provided flowchart illustrates a study that compares the classification performance between masked and non-masked adenoid image datasets (Fig. 1). The datasets undergo similar preprocessing steps, where images are cropped to remove any personal information and then resized to a standard size of 256 × 256 pixels. Data augmentation techniques are applied to both datasets to generate additional training samples, which helps improve the model’s robustness and generalization ability. The augmented datasets are subsequently divided into training, validation, and testing sets with an 80%, 10%, and 10% split, respectively.

Fig. 1.

The flowchart of classification of masked and non-masked adenoid images

The study employs an ensemble model to classify adenoid and non-adenoid images. It provides the combined strengths of multiple models to achieve better predictive performance. The model’s effectiveness is measured using several key metrics: accuracy, precision, recall, F1-score, and AUC [20]. These metrics provide a comprehensive evaluation of the model’s capability to accurately distinguish between adenoid and non-adenoid classes. By comparing the performance of masked versus non-masked datasets, the study aims to determine the impact of segmentation on classification results. This comparison is crucial as it explores whether preprocessing the images through segmentation enhances the model’s accuracy.

Material

Dataset

The provided dataset, sourced from Batman Training and Research Hospital, consists of 344 X-ray images that are focused on identifying the presence of adenoids in patients (Fig. 2). Adenoids are lymphatic tissues located in the back of the nasal cavity, which can sometimes become enlarged and cause breathing issues. Detecting this condition through medical imaging is crucial for diagnosis and treatment planning, which may involve surgical intervention.

Fig. 2.

The dataset of masked and non-masked adenoid images

To enhance the robustness of the dataset and overcome the limitations of its small size, data augmentation techniques were applied. Data augmentation is a common practice in medical imaging that generates modified versions of the original images through techniques such as flipping, rotation, and zooming [21, 22]. In this case, the original 344 images were expanded into 1216 images for training, 152 for validation, and 152 for testing. This step ensures that the deep learning model used for classification is exposed to a wider variety of data, and it improves its ability to generalize and recognize patterns that indicate the presence or absence of adenoids.

Segmentation plays a crucial role in this task, as shown in the two images provided: one being the original X-ray and the other showing the adenoid region highlighted in red. Segmentation helps isolate the area of interest, and it allows the model to focus on relevant anatomical features without interference from surrounding tissues. This not only improves the model’s learning process but also enhances the accuracy of its predictions.

Data Augmentation

The two images displayed represent a non-masked X-ray and a masked version, highlighting the importance of data preprocessing in medical image classification tasks (Fig. 3). The non-masked X-ray image shows the full anatomical structure of the head and upper spine, including background noise. When using this image for classification, the model would have to process all details, including irrelevant features like the background, which might reduce the accuracy and focus of the classification task.

Fig. 3.

The augmentation of masked and non-masked images

On the other hand, the masked X-ray image isolates a specific region of interest, highlighted in red, possibly focusing on a key area such as a bone or a detected abnormality. This type of image helps the model concentrate on the essential features relevant to the classification task, reducing the chances of noise affecting the performance. In medical imaging, where specific areas like fractures or diseases need to be analyzed, segmentation helps improve classification performance by eliminating distractions.

Masked images typically yield better classification results by removing unnecessary data. This allows the model to focus on critical areas, and it improves overall accuracy. In contrast, non-masked images may introduce background noise, leading to potential misclassifications. For better performance, data augmentation techniques such as rotation, scaling, and flipping can be applied to both sets of images. It helps the model generalize better by learning from a more diverse dataset. In conclusion, segmentation, combined with data augmentation, plays a critical role in improving the robustness and accuracy of medical image classification models.

Data Masking

The adenoid dataset is obtained from Batman Training and Research Hospital, and it required labeling for both the training and testing phases. To ensure accuracy in this process, we collaborated with two expert radiologists from the hospital’s Ear, Nose, and Throat (ENT) department. These radiologists, part of the research team, marked the ground truth of the adenoid images. The open-source image editing tool, GIMP, was used for marking the ground truth. The adenoids were highlighted in red to ensure visibility and consistency across the dataset. As evident from sample images, the shape, size, location, and orientation of the adenoid glands vary significantly, reflecting the complexity of the dataset (Fig. 4).

Fig. 4.

a Original image 1, b masked image 1, c original image 2, d masked image 2

Methodology

EfficientNet

EfficientNet is a scalable deep learning model designed to optimize accuracy and computational efficiency by uniformly scaling the depth, width, and resolution of the network. It utilizes compound scaling, which allows the model to maintain high accuracy while keeping computational demands low. EfficientNet has gained popularity for medical image classification tasks because of its balance between performance and resource efficiency.

MobileNet

MobileNet is a lightweight convolutional neural network designed for mobile and edge devices. Its architecture is based on depthwise separable convolutions which is reduced computational complexity and memory usage. MobileNet is particularly useful for applications where computational resources are constrained, without sacrificing significant accuracy. The second version, MobileNetV2, introduced features such as inverted residuals and linear bottlenecks that improve its ability to handle complex tasks like image classification.

ResNet50

ResNet (residual networks) [23] introduced the concept of residual learning to address the issue of vanishing gradients in deep networks. ResNet50 is a 50-layer deep model that uses identity mapping through skip connections which allowed it to train very deep networks efficiently. It is widely used for image classification and recognition tasks, which include medical imaging, due to its high accuracy and strong generalization capabilities.

ResNet152

ResNet152 [24] is a deeper variant of ResNet50 with 152 layers. It offers improved accuracy on large-scale image classification tasks by further exploiting residual learning. This model can handle more complex tasks but comes with higher computational costs. ResNet152 is often used when high precision is required, and computational resources are sufficient to support deeper networks.

VGG16

VGG16 [25] is a classic convolutional neural network that consists of 16 layers and is known for its simplicity and depth. Despite being computationally expensive compared to modern architectures, it has remained a popular choice for image classification due to its consistent performance. VGG16 is particularly effective in transfer learning scenarios which makes it a strong contender in various image classification benchmarks.

Xception

Xception [26] is an extension of the Inception architecture, which replaces the standard Inception modules with depthwise separable convolutions. This model emphasizes efficiency in training and evaluation while maintaining high accuracy. Xception is particularly useful in image classification tasks that require both accuracy and reduced computational overhead, although it may not perform as well as other architectures like ResNet50 on some datasets.

Evaluation Metrics

Evaluation metrics such as accuracy, precision, recall, and F1-score are essential for assessing the performance of classification models. The mathematical expression of these metrics is shown below:

$$\text{Accuracy}=\frac{TP+TN}{TP+FP+FN+TN}$$ 1

$$\text{Precision}=\frac{\mathit{TP}}{TP+FP}$$ 2

$$\text{Recall}=\frac{\mathit{TP}}{TP+FN}$$ 3

$$\text{F1-score}=\text{2 x}\frac{\text{Precision x Recall}}{\text{Precision x Recall}}$$ 4

TN, FP, TP, and FN denote true negative, false positive, true positive, and false negative of a confusion matrix, respectively.

Experiments and Results

The table outlines the specific parameters used for compiling and training six different deep learning models—EfficientNet, MobileNet, ResNet50, ResNet152, VGG16, and Xception—while comparing their performance on masked versus non-masked images (Table 1). For the compilation process, all models are set to use the Adam optimizer with a learning rate of 0.001, indicating a preference for this optimizer’s adaptive learning rate capabilities and effectiveness in handling sparse gradients. The input image size is standardized at 256 × 256 pixels for all models which is crucial for fair comparison. This consistent image size allows the models to learn from a fixed resolution, reducing variability due to different input dimensions. The chosen loss function, binary cross-entropy, is applied across all models that make it evident that the problem being addressed is a binary classification or segmentation task, where the output is evaluated against two possible classes.

Table 1.

The hyperparameters performance of deep learning models

Model	Masked images/non-masked images
	Image size	Epochs	Batch size	Optimizer	Learning rate	Loss
EfficientNet	256 × 256	25	16	Adam	0.001	BinaryCrossentropy
MobileNet	256 × 256	25	16	Adam	0.001	BinaryCrossentropy
ResNet50	256 × 256	25	16	Adam	0.001	BinaryCrossentropy
ResNet152	256 × 256	25	16	Adam	0.001	BinaryCrossentropy
VGG16	256 × 256	25	16	Adam	0.001	BinaryCrossentropy
Xception	256 × 256	25	16	Adam	0.001	BinaryCrossentropy

In terms of training, each model is trained for 25 epochs with a batch size of 16. The number of epochs suggests a balanced approach to training. A batch size of 16 is a moderate choice, balancing between efficient computation and stable gradient updates. By using identical compile and training parameters for both types of datasets, the study aims to isolate the impact of image segmentation on model efficacy. This setup is crucial for understanding whether the added complexity of masked images provides a tangible benefit in terms of model accuracy and robustness.

Performance on Masked Adenoid Images

The analysis of the table reveals significant differences in the performance of various models on masked images. ResNet50 emerges as the top performer that achieves perfect scores across all metrics and that indicates its outstanding capability to detect true positives while avoiding false positives (Table 2). Similarly, EfficientNet and VGG16 show strong performance, with an accuracy of 0.99 and near-perfect precision and recall values which makes them highly reliable models for this dataset.

Table 2.

The result of classification on masked adenoid images

Models	Accuracy	Precision	Recall	F1-score	AUC
EfficientNet	0.99	0.99	0.99	0.99	0.99
MobileNet	0.88	0.90	0.89	0.88	0.89
ResNet50	1.00	1.00	1.00	1.00	1.00
ResNet152	0.89	0.83	0.80	0.89	0.89
VGG16	0.98	0.99	0.99	0.99	0.99
Xception	0.75	0.65	0.61	0.59	0.61

In contrast, Xception performs the weakest, with an accuracy of 0.75 and an F1-score of 0.59. The low precision and recall values for Xception suggest difficulties in balancing these key metrics which indicate that this model may not be well-suited for the task or dataset used. On the other hand, MobileNet and ResNet152 show moderate performance, with accuracies of 0.88 and 0.89, respectively. However, ResNet152 has lower precision (0.83) and recall (0.80) which reveals a trade-off where the model performs better in avoiding false positives but misses more true positives compared to other models like ResNet50 and EfficientNet.

ResNet50 stands out as the most reliable model, while EfficientNet and VGG16 also show promising results. Xception, however, may need further optimization or may not be suitable for this segmentation task, and MobileNet and ResNet152 could benefit from tuning to achieve a better balance between precision and recall.

Performance on Non-masked Adenoid Images

The analysis of the table reveals significant differences in the performance of various models on non-masked images (Table 3). EfficientNet stands out as the top performer, achieving the highest accuracy (0.92), precision (0.93), and F1-score (0.92), indicating that it excels in both identifying true positives and minimizing false positives. This makes it a highly reliable model for working with non-masked image data.

Table 3.

The result of classification on non-masked adenoid images

Models	Accuracy	Precision	Recall	F1-score	AUC
EfficientNet	0.92	0.93	0.87	0.92	0.92
MobileNet	0.77	0.80	0.78	0.77	0.77
ResNet50	0.84	0.87	0.84	0.86	0.84
ResNet152	0.83	0.84	0.83	0.84	0.84
VGG16	0.82	0.83	0.82	0.81	0.82
Xception	0.71	0.62	0.70	0.65	0.62

In contrast, Xception performs the weakest, with an accuracy of 0.71 and an F1-score of 0.65. Its low precision (0.62) and recall (0.70) values suggest difficulty in balancing these metrics, indicating that this model may not be suitable for this task. MobileNet also shows relatively lower performance, with an accuracy of 0.77 and an F1-score of 0.77, suggesting that it struggles to handle non-masked images as effectively as the other models.

ResNet50 and ResNet152 deliver moderate performance, with accuracies of 0.84 and 0.83, respectively. While ResNet50 has higher precision (0.87) and a better F1-score (0.86), ResNet152 shows slightly lower precision (0.84) and F1-score (0.84), indicating that both models are well-balanced but slightly less effective than EfficientNet. VGG16 also shows moderate results, with an accuracy of 0.82 and F1-score of 0.81, performing comparably to ResNet but slightly behind in overall effectiveness.

EfficientNet proves to be the most effective model for non-masked image classification, while ResNet50 and ResNet152 offer good performance, though slightly less optimized. VGG16 performs moderately well, but MobileNet and Xception struggle with this dataset, particularly Xception, which may require further optimization to improve its performance.

This table compares various studies on adenoid detection, focusing on different models and datasets (Table 4). Zhao et al. [16] utilized a CNN based on Fujioka’s method for lateral cephalograms, achieving 91% accuracy and 90% recall. Dong et al. [18] used HMSAU-Net and 3D-ResNet on a CBCT dataset, also reaching 91% accuracy with a higher recall of 97%. Fu et al. [19] employed AFFMNet on patch-level and whole slide images, achieving 97% accuracy, though other metrics were not specified. Alshbishiri et al. applied U-Net to X-ray images of adenoid glands, reporting a Dice coefficient of 0.74 and an IoU of 0.62. He et al. used ADNet on MRI images, achieving a recall of 96% within a 15-pixel range. In comparison, our study using ensemble deep learning models on X-ray images from Batman Training and Research Hospital achieved perfect performance with ResNet50, boasting 100% accuracy, and EfficientNet and VGG16 showed 99% precision and recall, highlighting the superiority of ensemble models, especially on masked images.

Table 4.

The comparison of the proposed model and ensemble deep learning models

Study	Method	Dataset	Accuracy	Precision	Recall	F1-score	Dice	IoU
Zhao et al. (2021)	CNN based on Fujioka’s method	Lateral cephalograms of adenoids	0.91	_	0.90	_	_	_
Dong et al. (2023)	HMSAU-Net and 3D-ResNet	Cone-beam computed tomography (CBCT) dataset	0.91	_	0.97	_	_	_
Fu et al. (2023)	AFFMNet model	Patch-level and whole slide images	0.97	_	_	_	_	_
Alshbishiri et al	U-Net	X-ray images of adenoid glands	_	_	_	_	0.74	0.62
He et al	ADNet model	MRI images	_	_	0.96 within 15 pixels	_	_	_
Our model	Ensemble deep learning models	Batman Training and Research Hospital, 344 X-ray images	1.00 (ResNet50 on masked images)	99% (EfficientNet, VGG16)	99% (EfficientNet, VGG16)	99–100% (ResNet50)	_	_

Among the studies listed, ResNet50 from the current study stands out as the best-performing model, achieving 100% accuracy, along with 99–100% precision, recall, and F1-score on masked images. This exceptional performance demonstrates its effectiveness in handling the adenoid classification.

In contrast, Alshbishiri et al.’s U-Net model, while useful for segmenting adenoid glands, showed the poorest performance with a Dice coefficient of 0.74 and an IoU of 0.62. These lower metrics indicate that the model struggles to achieve the same level of segmentation accuracy and localization as the other models.

The comparison highlights the superiority of deeper ensemble models like ResNet50 for classification tasks, while simpler architectures like U-Net, though effective for segmentation, may fall short in accuracy and overall performance for similar tasks (Figs. 5, 6, 7, 8, 9, 10).

Fig. 5.

The training accuracy of deep learning models with non-masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Fig. 6.

The training accuracy of deep learning models with masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Fig. 7.

The ROC Curve of deep learning models with non-masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Fig. 8.

The ROC curve of deep learning models with masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Fig. 9.

The confusion matrix of deep learning models with non-masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Fig. 10.

The confusion matrix of deep learning models with masked images—a EfficientNet, b MobileNet, c ResNet50, d ResNet152, e VGG16, f Xception

Discussion

The findings of this study offer significant insights into the role of deep learning in adenoid classification using X-ray images with a particular focus on comparing masked and non-masked datasets. The results demonstrate that deep learning models can be highly effective in medical image analysis for adenoid classification. ResNet50, EfficientNet, and VGG16 emerged as the top-performing models on masked images, with ResNet50 achieving a perfect classification accuracy of 100%. This highlights the robustness and reliability of deep learning techniques in handling complex medical imaging tasks.

One of the key contributions of this research is the comparative analysis between masked and non-masked datasets. The results clearly indicate that image segmentation plays a critical role in improving the accuracy of adenoid classification. The models trained on masked images generally outperformed those trained on non-masked images across all evaluation metrics, including accuracy, precision, recall, F1-score, and AUC. This underscores the importance of preprocessing steps such as segmentation in enhancing the focus on the relevant anatomical regions, which allow the models to avoid being distracted by irrelevant background information.

However, the study also reveals that certain models, such as EfficientNet, can achieve relatively high accuracy even on non-masked images, with an accuracy of 92%. This suggests that while segmentation significantly improves model performance, certain architectures may still generalize well on raw images. This finding is particularly relevant in real-world clinical settings where segmentation may not always be feasible due to time or resource constraints. The fact that models like EfficientNet maintain strong performance without segmentation opens possibilities for deploying AI-driven diagnostic tools in resource-limited environments.

On the other hand, models like Xception performed poorly on both masked and non-masked datasets, with F1-scores as low as 59% on masked images and 65% on non-masked images. This points to the fact that not all deep learning architectures are equally suited for this task, and model selection plays a critical role in achieving optimal results. The lower performance of Xception also highlights the variability in how different models handle medical images, which may be influenced by factors such as the complexity of the architecture or the specific characteristics of the dataset.

Conclusion

In conclusion, this study demonstrates the effectiveness of ensemble deep learning models in the classification of adenoid hypertrophy using X-ray images. By evaluating both masked and non-masked datasets, the research highlights the significant impact of image preprocessing on model performance. Segmentation was shown to improve classification accuracy across all models. ResNet50 achieved a perfect score when applied to masked images. This underscores the critical importance of directing the model’s attention to the relevant anatomical region for optimal performance.

Despite the superior performance of masked images, models like EfficientNet were able to maintain high accuracy even on non-masked datasets, suggesting that some architectures are more resilient to noise and irrelevant features in medical images. This finding offers flexibility in clinical applications where segmentation may not be feasible.

The comparative analysis of different models also reveals that not all CNN architectures are equally suited for this task. While ResNet50, EfficientNet, and VGG16 performed well, Xception struggled to deliver competitive results, and it emphasizes the importance of model selection in medical image classification tasks.

Future research could explore further optimizations of the models, particularly in transfer learning, to enhance performance on both masked and non-masked datasets. Additionally, the integration of other preprocessing techniques or multimodal datasets could further improve the generalization capabilities of the models. Ultimately, this research contributes valuable insights into the growing body of work on AI-driven medical diagnostics, and it offers potential pathways for more accurate, reliable, and accessible solutions for adenoid classification.

Data Availability

The dataset generated and analyzed during this study is available from the corresponding author on reasonable request.

Declarations

Competing Interests

The authors declare no competing interests.