Purpose or Case Report: In this study, we trained two convolutional neural networks to automatically identify the third vertebral level and segment the abdominal muscle in contrast enhanced abdominal CT images. In the future, these models will be used to determine reference ranges for skeletal muscle mass in children by age for the purpose of identifying patient characteristics associated with differences in skeletal muscle mass.
Methods & Materials: 370 CT images of abdomen with contrast were randomly selected from 10,000 scans, corresponding to roughly 10 for each gender from ages 0-18. The third vertebral level was identified in the sagittal image series and the abdominal muscle was manually segmented in the axial series corresponding to the identified level. A convolutional neural network with the U-Net architecture was used to train the segmentation model. A 5-fold cross validation set was generated by randomly splitting the axial images into datasets with 70% training, 10% training validation, and 20% test images. A second dataset with a similar split was generated from augmented data created by flipping, rotating, shifting, zooming, and shearing the original input axial images resulting in a set of 5,550 images. These two sets of images were then preprocessed by algorithmically removing the scan table, thresholding, resizing, and normalizing the images. The segmentation models were trained until there was no improvement in the validation loss function for 20 consecutive epochs. The models’ performance was evaluated using the dice similarity coefficient (DSC). The performance of two different loss functions, a weighted dice score and binary cross entropy, were compared.
Results: The model trained with the non-augmented dataset achieved mean DSCs of 0.914 and 0.917 using the weighted dice score and binary cross entropy loss functions respectively. The model trained using the augmented dataset achieved means of 0.924 and 0.925.
Conclusions: The U-Net models trained on our datasets accurately segment the abdominal muscle on pediatric axial CT scans across the range of ages 0-18 years. Additional work is being performed to validate the performance of a U-Net model trained to identify the third vertebral level. Pending validation of this second model, an efficient pipeline for the segmentation of CT scans for the purpose of creating skeletal muscle nomograms has been developed.