Quantification of Thoracic Volume and Spinal Length of Pediatric Scoliosis Patients on Chest MRI Using a 3D U-Net Segmentation

Abstract

Background/Objectives: Adolescent idiopathic scoliosis (AIS) can lead to significant chest deformations. The quantification of chest deformity and spinal length could provide additional insights for monitoring during follow-up and treatment. This study proposes a 3D U-Net convolutional neural network (CNN) for automatic thoracic and spinal segmentations of chest MRI scans. Methods: In this proof-of-concept study, axial chest MRI scans from 19 girls aged 8–10 years at risk for AIS development and 19 asymptomatic young adults were acquired (n = 38). The thoracic volume and spine were manually segmented as the ground truth (GT). A 3D U-Net CNN was trained on 31 MRI scans. The seven remaining MRI scans were used for validation, reported by the Dice similarity coefficient (DSC), the Hausdorff distance (HD), precision, and recall. From these segmentations, the thoracic volume and 3D spinal length were calculated. Results: Automatic chest segmentation was possible for all chest MRIs. For the chest volume segmentations, the average DSC was 0.91, HD was 51.89, precision was 0.90, and recall 0.99. For the spinal segmentation, the average DSC was 0.85, HD was 25.98, precision was 0.74, and recall 0.99. Chest volumes and 3D spinal lengths differed by on average 11% and 12% between automatic and GT, respectively. Qualitative analysis showed agreement between the automatic and manual segmentations in most cases. Conclusions: The proposed 3D U-Net CNN shows a high accuracy and good predictions in terms of HD, DSC, precision, and recall. This suggested 3D U-Net CNN could potentially be used to monitor the progression of chest deformation in scoliosis patients in a radiation-free manner. Improvement can be made by training the 3D U-net with more data and improving the GT data.