Age and Sex Estimation in Children and Young Adults Using Panoramic Radiographs with Convolutional Neural Networks

Abstract

Image processing with artificial intelligence has shown significant promise in various medical imaging applications. The present study aims to evaluate the performance of 16 different convolutional neural networks (CNNs) in predicting age and gender from panoramic radiographs in children and young adults. The networks tested included DarkNet-19, DarkNet-53, Inception-ResNet-v2, VGG-19, DenseNet-201, ResNet-50, GoogLeNet, VGG-16, SqueezeNet, ResNet-101, ResNet-18, ShuffleNet, MobileNet-v2, NasNet-Mobile, AlexNet, and Xception. These networks were trained on a dataset of 7336 radiographs from individuals aged between 5 and 21. Age and gender estimation accuracy and mean absolute age prediction errors were evaluated on 340 radiographs. Statistical analyses were conducted using Shapiro–Wilk, one-way ANOVA, and Tukey tests (p < 0.05). The gender prediction accuracy and the mean absolute age prediction error were, respectively, 87.94% and 0.582 for DarkNet-53, 86.18% and 0.427 for DarkNet-19, 84.71% and 0.703 for GoogLeNet, 81.76% and 0.756 for DenseNet-201, 81.76% and 1.115 for ResNet-18, 80.88% and 0.650 for VGG-19, 79.41% and 0.988 for SqueezeNet, 79.12% and 0.682 for Inception-Resnet-v2, 78.24% and 0.747 for ResNet-50, 77.35% and 1.047 for VGG-16, 76.47% and 1.109 for Xception, 75.88% and 0.977 for ResNet-101, 73.24% and 0.894 for ShuffleNet, 72.35% and 1.206 for AlexNet, 71.18% and 1.094 for NasNet-Mobile, and 62.94% and 1.327 for MobileNet-v2. No statistical difference in age prediction performance was found between DarkNet-19 and DarkNet-53, which demonstrated the most successful age estimation results. Despite these promising results, all tested CNNs performed below 90% accuracy and were not deemed suitable for clinical use. Future studies should continue with more-advanced networks and larger datasets.

1. Introduction

When estimating age, criteria such as weight, height, signs of puberty, mental development, and teeth and bone development are considered [1]. Estimated age is essential in some dental treatment approaches. It can also be used in anthropology to determine the age of death of hominid fossils. In forensic cases, it can be used when chronological age information is unavailable or the existing information is questionable [2].

Hard tissues in the body can be examined to determine age. For bone-based estimation, the hand–wrist area, pubic symphysis, and medial clavicular epiphysis cartilage can be examined, but teeth-based methods are generally less susceptible to error [3]. Teeth are the most durable part of the human body and can provide valuable information about a person’s ethnic origin, occupation, sex, and age [4]. Cameriere’s, Demirjian’s, and Willems’s methods and other methods, such as measuring pulp volume or root transparency, have been developed to estimate age from teeth [5,6]. Each method has different levels of accuracy and is generally limited to specific populations [7]. Performing age estimation based on manual measurement methods such as Cameriere’s open apices method or using dental development tables like Demirjian’s method can increase the likelihood of errors and result in subjective outcomes that may vary between operators. Moreover, these conventional methods are time-consuming and require the teeth used for age estimation to be free of cavities or to have not undergone root canal treatment [8]. Indeed, it is essential to note that sex can influence dental development patterns. It has been demonstrated that girls reach certain stages of dental development earlier than boys, particularly in permanent teeth [9]. Therefore, age estimation methods such as those of Demirjian and Willems have separate dental development tables for girls and boys to account for these sex-related differences [10].

CBCT and cephalometric and panoramic radiography can be used to estimate radiographic sex based on teeth and jawbones [11,12,13]. Although panoramic radiographs have disadvantages, such as ghost images, geometric distortion, and superimpositions, they are frequently preferred for diagnosis and follow-up because they are cheaper and more accessible, can be taken quickly, and have a lower radiation dose [14].

Parallel to technological advancements and the integration of deep learning (DL) into various industries, traditional techniques are being replaced by artificial intelligence (AI) methods. DL is a subset of machine learning that uses networks with multiple layers of computation to analyze input data, thereby significantly enhancing the performance of AI [15]. Convolutional neural networks (CNNs) are a type of DL algorithm particularly well suited for image processing [16]. AlexNet is a classic CNN architecture with an eight-layer depth that won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012 by significantly surpassing previous methods. It became popular and inspired the development of other similar methods. VGG-16 is another popular method with a straightforward design, but it has many parameters that take a long time to learn. The numbers at the end of the network names usually indicate the number of layers, in other words, their depth. VGG-19 is very similar to VGG-16 but contains more layers. Google implemented inception building blocks into CNNs and called the network GoogLeNet (Inceptionv1). These blocks work well together, resulting in a model that is easy to generalize. This model is regularly updated with new versions, including Xception, which uses depthwise separable convolutions. ResNet, short for Residual Network, is a specific type of neural network. It comes in various sizes and numbers of layers, such as ResNet-18, ResNet-50, and ResNet-101. DenseNet is very similar to ResNet but concatenates the previous layer’s output with the next layer’s output. DarkNet-19 and DarkNet-53 are the networks used in YOLO, a popular object recognition method. CNNs generally require very high processing power. As alternatives to other models, SqueezeNet, ShuffleNet, and NasNet-Mobile were developed for use in less-powerful systems [17].

The study hypothesizes that at least one of the tested CNNs will have an accuracy greater than 90%, which is considered the success threshold. Our aim in this exploratory study is to compare the performance of 16 popular CNNs in predicting age and gender from panoramic radiographs in children and young adults.

2. Materials and Methods

This study was conducted according to the Declaration of Helsinki and approved by the Karaman Provincial Health Directorate and the Karamanoğlu Mehmetbey University Faculty of Medicine Clinical Research Ethics Committee with decision number 07-2022/19. Patient consent was obtained before the radiographs were taken and the consent forms have been archived at institutions.

In machine learning processes, the prediction accuracy of trained models is closely related to the sample size, with accuracy tending to improve as the sample size increases. However, after reaching a certain point, often referred to as the “sweet spot”, further increases in sample size do not significantly enhance the model’s performance. Identifying this point is a state-of-the-art research topic in biostatistics, and unfortunately, there is currently no generally accepted rule. Since having a large sample size does not harm the artificial intelligence training process aside from extending the required training time, the approach of using “the largest possible sample size” was preferred in this research. The dataset consisted of 7667 panoramic radiographs obtained from patients aged 5–21, who did not have any systemic diseases; these radiographs were previously taken for diagnostic or treatment purposes at Karamanoğlu Mehmetbey University’s Ahmet Keleşoğlu Faculty of Dentistry and Karaman Oral and Dental Health Center between 2016 and 2023. Out of a total of 8000 panoramic radiographs obtained using PCH-2500 (60–75 kVp, 5–10 mA, Vatech, Gyeonggi-do, Republic of Korea), Veraview IC5 HD (60–70 kV, 1–7.5 mA, Morita Corporation, Osaka, Japan), Hyperion X9 (60–75 kVp, 5–10 mA, My Ray, Imola, Italy), and Hyperion X5 (60–75 kVp, 5–10 mA, My Ray, Imola, Italy) panoramic radiography devices, 333 radiographs with imaging errors such as significant distortion, imaging errors due to patient positioning, or large ghost images were excluded from the study. To reflect reality more accurately, radiographs from patients with trans-palatal arches, braces, root canal treatments, caries, implants, bridges, and all types of restorations, hypodontia, or oligodontia were not excluded from the study. If present, metadata such as the name, date of birth, and radiograph capture date were cropped from specific radiographs.

When building a DL system, a significant focus must be on avoiding overfitting. The learning capacity must be appropriate, considering the amount of data and the difficulty of the task. Training, validation, and testing must be used to correctly verify which accuracy a DL system will achieve for future unseen data. This means that the acquired dataset must be split into three parts. The training set (1) is used to learn which patterns in the images correlate with the labels. For the validation set (2), once these patterns are learned, a different set of images is used to verify whether the model is under- or overfitting. Based on this assessment, the learning capacity of the network can be adapted to better fit the data. Regarding test set (3), since design choices are made based on the validation set’s performance, we need a third set to avoid selection bias and to report an unbiased estimate of the model performance. No perfect ratio can be determined a priori, but it should be determined based on the amount of available data. The basic rule is that the validation and test sets should be large enough to represent the whole range of variability found in real life. The remainder of the data can then be used for training. For a small amount of data, the ratio used for validation and testing will be relatively high compared to that used for training. However, as the amount of data grows, this ratio will become smaller [18]. Considering that we had a relatively medium-sized dataset, this study allocated 70% of our dataset for training, 25% for validation, and 5% (10 radiographs for each class) for testing. The numerical distribution of the radiographs according to age/sex and purpose of use is provided in

Table 1. The numerical distribution of the radiographs according to age/sex and purpose of use.

Age	Sex	Training	Validation	Test	Total
5	Male	178	75	10	263
	Female	164	69	10	243
6	Male	165	70	10	245
	Female	172	73	10	255
7	Male	285	122	10	417
	Female	277	118	10	405
8	Male	176	75	10	261
	Female	260	111	10	381
9	Male	237	101	10	348
	Female	190	81	10	281
10	Male	157	67	10	234
	Female	166	71	10	247
11	Male	147	62	10	219
	Female	125	53	10	188
12	Male	110	47	10	167
	Female	119	50	10	179
13	Male	96	41	10	147
	Female	103	44	10	157
14	Male	90	38	10	138
	Female	106	45	10	161
15	Male	75	32	10	117
	Female	115	49	10	174
16	Male	77	33	10	120
	Female	129	55	10	194
17	Male	108	46	10	164
	Female	178	76	10	264
18	Male	108	46	10	164
	Female	180	76	10	266
19	Male	101	43	10	154
	Female	159	68	10	237
20	Male	113	48	10	171
	Female	180	77	10	267
21	Male	116	49	10	175
	Female	185	78	10	273
Total		5147	2189	340	7676

.

The chronological age predictions were conducted using CNNs. In this study, several popular CNN architectures were employed: DarkNet-19, DarkNet-53, Inception-ResNet-v2, VGG-19, DenseNet-201, ResNet-50, GoogLeNet, VGG-16, SqueezeNet, ResNet-101, ResNet-18, ShuffleNet, MobileNet-v2, NasNet-Mobile, AlexNet, and Xception. Different radiography devices produce images at varying resolutions, necessitating the standardization of radiographs for training each CNN model. Therefore, all radiographs were automatically resized by adhering to the input resolution specifications outlined in

Table 2. Age, sex, and combined prediction accuracy with mean absolute error (MAE) and its descriptives.

	Input Resolution	Age Prediction Accuracy	Sex Prediction Accuracy	Age and Sex Prediction Accuracy	n	Mean Absolute Age Error	Std. Dev.	Std. Err.	95% Confidence Interval		Min.	Max.
									Lower	Upper
DarkNet-19	256 × 256	80.59%	86.18%	70.59%	340	0.4265	1.18641	0.06434	0.2999	0.5530	0	10
DarkNet-53	256 × 256	77.35%	87.94%	70.88%	340	0.5824	1.55783	0.08449	0.4162	0.7485	0	14
Inception-Resnet-v2	299 × 299	68.53%	79.12%	59.41%	340	0.6824	1.49107	0.08086	0.5233	0.8414	0	12
VGG-19	224 × 224	67.94%	80.88%	59.71%	340	0.6500	1.18415	0.06422	0.5237	0.7763	0	8
DenseNet-201	224 × 224	67.06%	81.76%	59.71%	340	0.7559	1.30434	0.07074	0.6167	0.8950	0	7
ResNet-50	224 × 224	66.76%	78.24%	58.82%	340	0.7471	1.39769	0.07580	0.5980	0.8962	0	7
GoogLeNet	224 × 224	64.12%	84.71%	58.24%	340	0.7029	1.15086	0.06241	0.5802	0.8257	0	5
VGG-16	224 × 224	63.82%	77.35%	55.00%	340	1.0471	1.92273	0.10427	0.8420	1.2522	0	8
SqueezeNet	227 × 227	62.65%	79.41%	54.41%	340	0.9882	1.76324	0.09562	0.8001	1.1763	0	10
ResNet-101	224 × 224	60.29%	75.88%	51.18%	340	0.9765	1.71735	0.09314	0.7933	1.1597	0	9
ResNet-18	224 × 224	58.53%	81.76%	50.59%	340	1.1147	1.86611	0.10120	0.9156	1.3138	0	15
ShuffleNet	224 × 224	57.65%	73.24%	47.94%	340	0.8941	1.38916	0.07534	0.7459	1.0423	0	7
MobileNet-v2	224 × 224	51.47%	62.94%	42.06%	340	1.3265	2.05296	0.11134	1.1075	1.5455	0	12
NasNet-Mobile	224 × 224	49.12%	71.18%	40.29%	340	1.0941	1.66960	0.09055	0.9160	1.2722	0	12
AlexNet	227 × 227	47.35%	72.35%	37.06%	340	1.2059	1.66660	0.09038	1.0281	1.3837	0	9
Xception	299 × 299	44.12%	76.47%	35.29%	340	1.1088	1.54096	0.08357	0.9444	1.2732	0	9

.

The training parameters, such as batch size, learning rate, and optimization method, significantly impact the accuracy of CNN networks. The optimal combination of these parameters can vary depending on the dataset’s structure. Therefore, experiments were conducted using three randomly selected networks (SqueezeNet, GoogLeNet, and DarkNet-19) with 36 different parameter combinations to identify the optimal training parameters. These combinations included batch sizes of 8, 16, and 32; learning rates of 0.01, 0.001, and 0.0001; and optimization methods Adam, SGDM, and RMSProp. Each network was trained for 50 epochs.

It was observed that as the number of epochs increased during AI training, the validation loss also increased, indicating overfitting of the model. To prevent overfitting, the number of epochs was set to 10. For learning curve optimization, a learning rate of 0.001 was used for the first five epochs, followed by a learning rate of 0.0001 for the subsequent five epochs. The batch size was set to 16, and the optimization method chosen was SGDM. After determining these parameters, all networks were trained using these specified values on MatLab (R2023a) software.

The performance of the models in age determination was evaluated on 10 radiographs from each age–sex combination, totaling 340 radiographs. The test set comprised approximately 5% of our entire radiograph dataset. The predictive performance of the trained models was evaluated based on age, sex, and a combination of both. The accuracy percentage was calculated by dividing the correctly predicted radiographs by the total number of radiographs in the specified test group. We determined each network model’s mean absolute error (MAE) by calculating the absolute difference between predicted and chronological age for a more detailed age prediction performance evaluation.

Statistics

The MAE data were analyzed using Shapiro–Wilk’s normality test and Levene’s homogeneity test with the IBM SPSS Statistics (v. 21, IBM, Chicago, IL, USA) software package. For the normally and homogeneously distributed dataset, one-way analysis of variance (ANOVA) and post hoc Tukey tests were performed. A p-value of less than 0.05 was considered statistically significant.

3. Results

According to Table 2, DarkNet-19 achieved the highest accuracy in age prediction alone, with an accuracy of 80.59% and a mean absolute error (MAE) of 0.4265. On the other hand, DarkNet-53 achieved the highest accuracy in sex prediction alone, with an accuracy of 87.94%. The detailed prediction matrix for DarkNet-53 is shown in Figure 1.

Table S1 (Supplementary Materials) provides a detailed statistical comparison of the MAE of all networks in the study. Accordingly, the differences between DarkNet-19 and DarkNet-53, the networks that provide the closest results to chronological age, and AlexNet and Xception, which give the most inaccurate results, are not statistically significant. In contrast, the difference between DarkNet-19 and Xception is significant.

Additionally, when age and sex predictions were combined, the accuracy was lower than when they were studied separately. DarkNet-53 achieved the highest accuracy in combined prediction, 70.88%.

When examining the prediction success distribution graphs in Figure 1 and Figure 2, it can be observed that CNNs mostly made predictions close to the actual sex and age but sometimes made significantly incorrect predictions. For example, DarkNet-53 incorrectly predicted that one in ten radiographs of a 19-year-old female was that of a 5-year-old male (Figure 1). There were also cases where certain age groups were collectively mispredicted, such as GoogLeNet failing to make any correct predictions for the 17-year-old female group (Figure 2).

4. Discussion

AI is increasingly used in complex decision-making, problem-solving, and object recognition tasks. Studies evaluating the use of AI in age estimation have shown that AI can be more successful than manual techniques [19]. Thus, this study focused on AI instead of manual methods and aimed to evaluate the performance of 16 different convolutional neural networks (CNNs) in predicting age and gender from panoramic radiographs in children and young adults. The networks tested included DarkNet-19, DarkNet-53, Inception-ResNet-v2, VGG-19, DenseNet-201, ResNet-50, GoogLeNet, VGG-16, SqueezeNet, ResNet-101, ResNet-18, ShuffleNet, MobileNet-v2, NasNet-Mobile, AlexNet, and Xception. The hypothesis of the study that at least one artificial intelligence network would achieve an accuracy greater than 90% in predicting age and gender was rejected. The highest accuracy rate observed for age prediction was 80.59%, while the highest accuracy rate for gender prediction was 87.94%.

The use of AI in forensic medicine and dentistry has been explored using various CNNs, dental X-rays, MRIs, and photographs [20,21,22,23,24,25,26]. Panoramic radiographs are commonly used in routine diagnosis, treatment planning, and follow-up of treatments in the maxillofacial region. They are easily accessible imaging methods and are frequently preferred in age estimation studies using AI [20,23,24,27,28,29]. For these reasons, panoramic radiographs obtained from patients in the past were also used for this study.

An additional advantage of age estimation using AI, in addition to the shorter completion time compared to traditional methods [30], is that it can be used for age estimation even in cases where conventional methods are not applicable, such as when dealing with individuals who have dental caries, missing teeth, restorations, or root canal treatments. AI-based age estimation has no limitations in these cases [20]. As a result, the radiographs of individuals who had undergone dental treatments were not excluded from our study, and a considerable number of panoramic radiographs could be used in our analysis. This enabled a more comprehensive and inclusive approach to age estimation using AI, encompassing a diverse range of individuals, including those who had received dental treatments. In addition to the chosen network, the most crucial factors for the success of artificial intelligence models are the data sample size, data heterogeneity, and training epochs [31]. Considering that the replacement of primary teeth with permanent teeth and the development of jawbones occur dramatically between the ages of 5 and 21, patients within this age range were selected for the study to provide precise data for the AI. Furthermore, some laws differ depending on age, with ages 14, 18, and 21 generally considered to be legal thresholds [32].

Ataş et al. [24] used 1332 panoramic radiographs obtained from patients who were aged from 8 to 68 for their study, and they reported the mean absolute error (MAE) values as follows: 3.47 for EfficientNetB4, 4.35 for ResNet50V2, 4.22 for DenseNet-201, 3.44 for InceptionV3, 3.68 for MobileNetV2, and 10.46 for VGG16. They also modified the InceptionV3 architecture and achieved an MAE of 3.13. We obtained lower MAE values in our study group aged 5 to 21. The MAE values were 0.75 for DenseNet201, 1.32 for MobileNetV2, and 1.04 for VGG16. It is well known that input size has a significant impact on the success of machine learning. In our study, the use of a larger input size, along with our focus on a younger age range, may explain the observed difference.

According to the study conducted by Mualla et al. [23] using 1429 panoramic radiographs and employing AlexNet and ResNet-101, they reported an accuracy of 95.80% for AlexNet and 92.60% for ResNet. In our study, however, we obtained lower accuracies of 60.29% for ResNet-101 and 47.35% for AlexNet. It is important to note that, unlike in their research, we did not apply any additional pre-processing to the images, which may have contributed to the differences in the results.

In the study conducted by Kim et al. [27], they used cropped images of teeth from panoramic radiographs, specifically teeth numbers 16, 26, 36, and 46, and employed the ResNet-152 network for dental age estimation. They reported an accuracy range between 89.05% and 90.27% in their findings. Similar studies can be found in the literature where cropped panoramic images were used [28]. Using a deeper network and providing a smaller region of interest instead of a broad panoramic view for AI training may have contributed to higher performance.

Ko et al. [13] used 15,000 panoramic radiographs and Darknet-19, and they reported an accuracy of approximately 84% and 96% when the acceptable range of dental estimation was ±5 and ±10 years, respectively. They also reported a mean error between actual age and predicted age of 2.74 and 3.49, respectively. In our study, using DarkNet-19, which showed the highest accuracy in age estimation, we achieved an accuracy of 80.59% and an MAE of 0.42 ± 1.18641. We believe that the more significant number of panoramic radiographs used in Ko et al.’s study may have contributed to a higher accuracy than in ours.

Although there is no definitive consensus in the literature, it is generally reported that at least one thousand examples per class are necessary for AI image classification. The most significant limitation of this study is that it was conducted with a dataset that cannot be considered large due to the inherent constraints of the healthcare field. Another limitation worth mentioning is that the images needed to be pre-processed. Similar studies that included pre-processing have observed higher success rates. Future studies could utilize larger datasets with pre-processed images. Additionally, instead of the conventional CNNs we tested, networks specifically designed for medical imaging could be developed.

5. Conclusions

Although artificial intelligence offers promising results in estimating age and gender from dental panoramic radiographs, it has been concluded that none of the examined networks are suitable for clinical use without any pre-processing on panoramic radiographs or any additional algorithms on networks, since their accuracy remains below 90%. Artificial intelligence technology is developing very rapidly, and in the future, studies should be continued with more-advanced networks using a larger dataset.