DENSEN: a convolutional neural network for estimating chronological ages from panoramic radiographs

Abstract

Background: Age estimation from panoramic radiographs is a fundamental task in forensic sciences. Previous age assessment studies mainly focused on juvenile rather than elderly populations (> 25 years old). Most proposed studies were statistical or scoring-based, requiring wet-lab experiments and professional skills, and suffering from low reliability.

Result: Based on Soft Stagewise Regression Network (SSR-Net), we developed DENSEN to estimate the chronological age for both juvenile and older adults, based on their orthopantomograms (OPTs, also known as orthopantomographs, pantomograms, or panoramic radiographs). We collected 1903 clinical panoramic radiographs of individuals between 3 and 85 years old to train and validate the model. We evaluated the model by the mean absolute error (MAE) between the estimated age and ground truth. For different age groups, 3-11 (children), 12-18 (teens), 19-25 (young adults), and 25+ (adults), DENSEN produced MAEs as 0.6885, 0.7615, 1.3502, and 2.8770, respectively. Our results imply that the model works in situations where genders are unknown. Moreover, DENSEN has lower errors for the adult group (> 25 years) than other methods. The proposed model is memory compact, consuming about 1.0 MB of memory overhead.

Conclusions: We introduced a novel deep learning approach DENSEN to estimate a subject's age from a panoramic radiograph for the first time. Our approach required less laboratory work compared with existing methods. The package we developed is an open-source tool and applies to all different age groups.

Keywords: Chronological age estimation; Forensic anthropology; Orthopantomogram; Soft Stagewise Regression Network.

Fig. 1

Performance of DENSEN in children, teens, young adults and adults groups. a–d display the performance of the DENSEN for the four age groups. The figures show that the predictor works ideally in children and teens groups with MAEs of 0.6885 and 0.7615, respectively. MAEs increase in the young adult and adult groups to MAE 1.3502 and 2.8770, respectively

Fig. 2

Performance of Bayesian CNN in children, teens, young adults and adults groups. a–d show the performance of Bayesian CNN Net in these four groups. a suggests that there is generally a high correlation between the ground truth ages and predicted ones in the children group with MAE 0.5847. In the teen group, there is a slight increase in MAE (MAE = 0.8834). c and d present the model with difficulty handling the age estimation in young adults (MAE = 1.7232) and adult (MAE = 6.7267) groups

Fig. 3

Performance of DANet in children, teens, young adults and adults groups. a–d present DANet performance in the four age groups. The MAEs generated by the DANet are 0.5208, 0.7105, 2.0225, and 4.5547 for children, teens, young adults, and adult groups, respectively. The results display that the model estimates age accurately in the children group and teen group. However, in the young adult and adult groups, the model works with unfitted MAE 2.0225 and 4.5547, respectively

Fig. 4

Saliency maps of children, teens, young adults and adults groups. We randomly chose images within the four groups. Then we feed the images into DENSEN. After calculating the last linear layer’s gradient, we feed the photos into a previous convolutional layer to visualize the saliency map’s learned features

Fig. 5

Number of participants according to ages in train set (a) and test set (b)

Fig. 6

Sliding augmentation

Fig. 7

Structure of DENSEN. a shows that the network adopts a 2-stream model similar to its initial network architecture. There are two heterogeneous streams. For both streams, the basic building block is composed of convolution layers, batch normalization, non-linear activation, and pooling layers. b presents different activation functions (ReLU versus Tanh) and pooling (average versus maximum) adopted for each stream