In this study, we used deep learning model to estimate the age of children based on the MR signal changes associated with myelination process on T1 and T2-weighed images. Brain MR images of 119 children age ranging from 0.25 to 24 months were first used as a training and test dataset. The age was then estimated by deep learning model based on the T1-WI and T2-WI dataset and T1-WI only dataset. Our results showed that convolution neural network model using T1WI and T2WI dataset demonstrated higher correlation and lower mean absolute error (MAE) compared to T1-WI only dataset.
Methods
Data Acquisition and Pre-Processing
This study included 119 children age ranging from 0.25 to 24-month-old without any abnormalities in neurological and brain structural MRI examination (Figure 1). Whole brain 2D T1- and T2-WIs were obtained using a 1.5T (MAGNETOM Avanto, Siemens Healthcare, Erlangen, Germany; T1WI: TR/TE = 530/11 or 602/13, T2WI: TR/TE = 4040/92 or 4100/94) and a 3T (Achieva, Philips Healthcare, Amsterdam, Netherlands; T1WI: TR/TE = 2256/10 or 2155/10, T2WI: TR/TE = 4000/80 or 2000/140) MR scanners with 4.5-5.5 mm slice thickness. Axial images in 5 different levels: (a) middle cerebellar peduncle of the pons, (b)midbrain, (c) splenium of the corpus callosum and internal capsule, (d) centrum semiovale, and (e) subcortical white matter were extracted and adopted as a training and test dataset for ML. These images were converted from digital imaging and communications in medicine into bitmap then cropped and resized to a resolution of 128×128. Fig. 2 and 3 show the example of preprocessed images and the full data, respectively.
DL framework
Two types of deep learning neural network models were constructed using SONY neural network console ver. 1.20 (https://dl.sony.com/) on a Windows PC Intel Corei7 2.2GH, 32GB memory, with graphical processing unit NVIDIA GeForce GTX 1070. Each learning model consisted of 3 convolutional layers and 6 fully connected layers for T1WI only dataset and T1- and T2-WIs dataset (Figure 4). To build the ML model, the following solver parameters were used for training: 100 epochs; base learning rate for untrained model, Adam (learning rate = 0.001, beta_1 = 0.9, beta_2 =0.999, epsilon = 0.00000001).
Evaluation of the model
In evaluation of the deep learning model, the 4-fold cross-validation method was employed. For a validation, test data was evaluated 10 times and these output values were averaged to obtain final output. The correlation coefficient and Mean Absolute Error (MAE) were calculated between estimated age by ML model and true age using SPSS version 13.
Results
The scatterplots between estimated age by ML model and true age are shown in Figure 5. For the deep learning model using T1-WI only dataset, the correlation coefficient and MAE were 0.921 (p < 0.001) and 1.88 months, respectively. While for T1- and T2WIs dataset, the correlation coefficient and MAE were 0.931 (p < 0.001) and 1.79 months, respectively. Furthermore, by adding T2-WI to T1-WI dataset, the age overestimation in children below 12 months and age underestimation in children above 12 months were improved.Discussion
Brain myelination in children’s brain begins before birth and progresses rapidly within the first two years after birth started from the caudal to rostral, dorsum to ventral and central to peripheral parts of the brain.5-7 Myelination trajectories appear as white matter signal changes conventional T1- and T2-WIs reflecting the change of tissue water and myelin lipids.1-4 Barkovich et al. were then established a chart of age based on signal change on T1- and T2-WIs and this milestone chart helps radiologists who specialize in the neuroimaging of children.8-13 The ML is superior in recognizing shape, color and geographic pattern of an object.14 Therefore, it is expected that ML can recognize MR signal changes reflecting the myelination and can estimate the development of children’s brain. In our investigation, the estimated age provided by deep learning models showed strong correlations with true age. Adding T2-WI to T1-WI dataset improved the accuracy of the deep learning model in children below 12 months and above 12 months.Conclusion
A deep learning neural network model can estimate the progression of the children brain myelination based on MR signal distribution.Figure 1. Distribution of subjects based on the age.
All subjects included 119 children from 0.25 to 24-month-old after birth. They didn’t have any neurologic abnormality.
Figure 2. Examples of preprocessed dataset from one subject.
Axial images in 5 different levels consisting of (a) middle cerebellar peduncle of pons, (b) Midbrain, (c) splenium of corpus callosum and internal capsule, (d) centrum semiovale and (e) subcortical white matter were extracted and adopted as a training and test dataset for ML.
Figure 4. Deep learning neural network architecture which was used in this study.
Each learning model consisted of 3 convolutional layers and 6 fully connected layers for T1WI only dataset and T1- and T2-WIs dataset. To build the ML model, the following solver parameters were used for training: 100 epochs; base learning rate for untrained model, Adam (learning rate = 0.001, beta_1 = 0.9, beta_2 =0.999, epsilon = 0.00000001).
Figure 5. Scatterplots between estimated age by ML model and true age.
Red solid is y = x and blue dotted line is Trendline. For the deep learning model using T1-WI only dataset, the correlation coefficient and MAE were 0.921 (p < 0.001) and 1.88 months, respectively. While for T1- and T2WIs dataset, the correlation coefficient and MAE were 0.931 (p < 0.001) and 1.79 months, respectively. Furthermore, by adding T2-WI to T1-WI dataset, the age overestimation in children below 12 months and age underestimation in children above 12 months were improved.