Table 1.
Gender and age distribution of the dataset.
Fig 1.
Preprocessing to extract the femur and mandible voxels.
Initially, the femur and mandible parts were extracted from whole-body (head to toes) PMCT images, followed by segmentation of their corresponding bone parts.
Fig 2.
Mandible bone extraction from a raw slice image.
(a) Load the original image and crop its center with an image size of 256×256. (b) Apply a threshold of 400 to remove tissue and unrelated parts. (c) Perform morphological opening operation, a technique to remove small objects, while maintaining the shape and property of large objects. (d) Label components to identify objects. (e) Retain the large objects in labeled components. (f) Binary masked input to obtain the bone part of mandible slice.
Fig 3.
Overview of our age estimation architecture.
3D PMCT volumes of the femur and mandible are fed into the deep CNN, which then fuses their results to obtain the final precise prediction (years).
Fig 4.
3D deep CNN architectures for age assessment.
(a) Basic elements of the Resnet model, containing a convolutional layer with a number of filters (F), batch normalization (BN), and ReLU activation function; (b) bottleneck element of the Resnet model, which is an extension of the basic block by adding one more convolutional layer; (c) 3D CNN model based on Resnet 34, containing basic blocks with a different number of F; (d) 3D CNN model based on Resnet 50, including bottleneck blocks with a different number of F.
Table 2.
Age estimation results obtained based on the mandible and femur separately, and their corresponding fusion results using both male and female samples.
Table 3.
Age estimation results for mandible and femur alone, and their corresponding fusion results using only male samples.
Table 4.
Comparing age estimation results using SqueezeNet, MobileNet, and Resnet models for extracting femur features.
Fig 5.
Scatter plot showing the final prediction results achieved by fusing the mandible and femur.
This plot shows the correlation between the predicted and actual ages. The points follow a linear pattern, implying that there exists a high linear correlation. The increasing amount of green dots close to the linear line indicates the improving accuracy of the age estimation.