Development and Evaluation of Deep Learning-Based Automatic Segmentation Model for Skull Zero TE MRI in Children

Purpose: To develop and evaluate a deep learning technique to automatically segment bone structures in zero echo time (ZTE) for skull magnetic resonance imaging (MRI) in children. Materials and Methods: From January to December 2021, 38 bone ZTE MRIs from infants and children (age range, 1-31 months) were collected for model development. Mask images were generated by manually segmenting the craniofacial bone using a commercial segmentation program. Among them, 35 ZTE series were used to train the three-dimensional (3D)-nnUnet deep learning model and the remaining three series were used for model validation. A temporally different dataset of 19 ZTE bone MRIs obtained in May 2022 from infants and children (age range, 3-168 months) was used to determine the model's performance. Dice similarity coefficient was calculated for each test case. From 3D volume rendering images, segmentation accuracy, overall image quality, and visibility of cranial sutures were subjectively evaluated on a 5-point scale and compared with ground truth data from manual segmentation. Reasons for segmentation failure were analyzed using axially segmented ZTE images. Results: For the test set, the mean Dice similarity coefficient was 0.985 +/- 0.019. The segmentation accuracy was lower than the ground truth without showing a statistically significant difference between the two (3.39 +/- 1.11 vs. 3.73 +/- 0.77, p = 0.055). The overall image quality and suture visibility showed no significant difference (3.34 +/- 0.75 vs. 3.42 +/- 0.69, p = 0.317; 3.55 +/- 0.97 vs. 3.60 +/- 0.95, p = 0.157). Common reasons for low segmentation accuracy were well-pneumatized sinuses, metal artifacts, skin at the vertex level, and bones too thin. Conclusion: The deep learning-based automatic segmentation technique of bone ZTE MRIs showed comparable segmentation performance to manual segmentation. Using the deep learning-based segmentation results, acceptable 3D-volume rendering images of craniofacial bones were generated.