Using deep learning to estimate infant age from skull x-ray images

Data collection

Infants under 12 months of age who underwent plain skull radiography for head trauma evaluation between January 2010 and December 2021 were included in this study. Patients with congenital cranial deformities were excluded. This study was approved by the Institutional Review Board of Hallym University Sinseong Hospital (approval number 2023-01-002). As this is a retrospective study, informed consent was waived by the Institutional Review Board of Hallym University Sinseong Hospital (approval number 2023-01-002), and all imaging data was anonymized.

All skull plain X-ray images (74 kVp, 200 mA, 100 ms) were taken using a digital X-ray machine (GC85A, SAMSUNG, Korea) and were retrieved in DICOM format from the institution's picture archiving and communication system (PACS, Infinite version 3.0.9) and converted to .png format. To ensure patient confidentiality, personal information and annotations were removed during the conversion process, and images that were not properly focused were systematically excluded from the database. To maintain the integrity of the dataset, the acquired images were evaluated by a neurosurgery expert (HSL), and AP, Town, and lateral view radiographs that deviated significantly from the standard were specifically excluded.

Additionally, this study used another 864 images of 216 patients from the same institution for external validation. These images were obtained from skull radiographs performed between January 2017 and December 2021. The mean vertical resolution of the internal dataset was 2046 ± 80 (1382–2177) pixels, and the mean horizontal resolution was 1703 ± 74 (1382–2177) pixels. In the external validation dataset, the mean vertical resolution was 1991 ± 188 (1453–2177) pixels, and the mean horizontal resolution was 1641 ± 181 (1160–1814) pixels.

The skull X-ray images included four types of views: anterior-posterior (AP), Towne, right lateral, and left lateral. Some patients had all four types of X-rays, while others had only some of them. Before the study began, the X-ray images were categorized into two groups: the AP dataset (which included both AP and Towne views) and the lateral dataset (which included right and left lateral views). Two deep learning models were developed separately for these datasets.

Building the Dataset

Each image was labeled according to the patient's age group and classified into 12 categories by age in months. As shown in Table 1, the entire dataset was split into three subsets using random sampling in a ratio of 8:1:1: training dataset, validation dataset, and test dataset. These sub-datasets were mutually exclusive. The validation dataset was used to determine the optimal training process point. Sampling was performed stratified by age group to maintain a consistent data ratio in each subset. To increase the reliability of the performance, the dataset split was performed three times with three different seeds to train the deep learning model separately.

Data Preprocessing

To eliminate potential bias in age predictions unrelated to the skull, all images were preprocessed to hide teeth and sinus regions. Regions of exclusion (ROEs) were identified on the skull radiographs, including the orbit and mandibular regions. The ROE borders were defined as follows:

1. 1.

   On AP or Towne skull radiographs, they surrounded the upper border of the orbital rim and the lower border of the mandible (Fig. 1A).

2. 2.

   Lateral skull radiographs revealed lesions in the supraorbital rim, the anterior-most portion of the mandible, and the posterior edge of the cervical spinous process ( Fig. 1B ).

The defined regions in each of the 293 skull X-ray images were labeled as ROEs by a neurosurgery expert (HSL). The entire ROE dataset was split into training, validation, and test datasets by random sampling in a ratio of 8:1:1. A MobileNetV3 model was trained for object detection on the labeled ROEs. In terms of training parameters, an Adam optimizer was used with an initial learning rate of 1e − 3 and a batch size of 16. Afterwards, post-processing was performed on all images to remove the detected ROEs based on the following criteria:

1. 1)

   AP or Townes skull x-ray shows the area below the upper border of the ROE has been removed (Figure 1C)

2. 2)

   On the lateral skull radiograph, a rectangular box defined by the superior border of the ROE and the right border of the ROE was removed ( Fig. 1 D).

All aligned images were then reviewed by a neurosurgeon (HSL) and adjusted for any misprocessing. After aligning the regions of interest (ROIs) in the images, all images were zero-padded centrosymmetrically to make them square, matching their long sides in width and height. Bilinear interpolation was applied to the transformed square images of various sizes to resize them to a uniform size of 1024 × 1024 pixels. Min-max normalization was applied to normalize all images.

Training the CNN model

Two different CNN architectures, DenseNet-121 and EfficientNet-V2-M, were employed to build the deep learning model. DenseNet-121 has improved algorithms for feature representation and learning efficiency, and is effective for medical image classification.^Tenand EfficientNet-V2-M, a relatively recent example that has shown high performance in common image classification tasks at low computational cost.^11,12In brief, DenseNet consists of dense blocks that link feature maps from the previous layer, while the EfficientNet-V2-M model uses neural architecture search to find the most effective CNN architecture, similar to EfficientNet. DenseNet-121 and EfficientNet-V2-M were previously trained on the ImageNet dataset and fine-tuned by unboxing the weights.^11,12,13All layers are now unfrozen, allowing you to tweak all layers in the network.

The batch size was set to 8 for DenseNet-121 and 4 for EfficientNet-V2-M, which is the maximum that the hardware's GPU memory could handle for each architecture. Categorical cross-entropy was used as the loss function and the Adam optimizer was applied.¹⁴The initial learning rate was set to 0.0001 and decreased by 0.1 every 10 epochs. After the 20th epoch, early stopping was employed with a patience value of 10. The patience value counts successive training steps based on the loss or validation loss value on the tuning dataset to complete training within a total of 100 epochs. During training, if the validation loss value exceeded the minimum validation loss achieved so far in any epoch, the model was not saved. Therefore, the model updated in the epoch showing the minimum validation loss in the training process was selected as the model to be finally saved to prevent overfitting.

The deep learning models used in this study were implemented on the PyTorch platform using a hardware system consisting of an NVIDIA GeForce RTX 4090 graphics processing unit and an Intel Xeon Silver central processing unit with a customized liquid cooling system.

Performance evaluation and statistical analysis

After training the deep learning models, we evaluated the performance of each model three times on the test dataset with different seeds. For external validation, we tested the trained deep learning models on a separate external validation dataset as described above.

The primary outcome measure of the established deep learning model was the classification accuracy in predicting the 12 age groups separated by months. Secondary outcomes included the 1-month relaxation accuracy of the deep learning model. Continuous variables are presented as means with standard deviations. Mann–Whitney U tests were used to compare prediction performance between different age groups. P values ​​less than 0.05 were considered statistically different and all tests were two-sided. Gradient-weighted class activation maps (Grad-CAM++) were implemented in the neural network layer to identify discriminative regions that the deep learning tool uses to determine a specific class in a given image.¹⁵To verify the superiority of the method proposed in this study, a comparative experiment was conducted with the RSNA Bone Challenge Winner Model.¹⁶The RSNA Bone Challenge is a competition to estimate bone age based on hand x-rays of pediatric patients. The RSNA winning model not only used InceptionV3 as the deep learning network, but also used gender as an additional input feature.

Ethical approval

All procedures performed in studies involving human subjects were in accordance with the ethical standards of the institutional and/or national research committee (Institutional Review Board of Hulme University Holy Hospital) and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

This study was conducted as a retrospective analysis, and all patient data were anonymized before use. Due to the retrospective nature of the study, informed consent was exempted by the Ethics Committee of Hallym University Holy See Hospital (No. 2023-01-002).