Vertebra Fracture Classification from 3D CT Lumbar Spine Segmentation Masks Using a Convolutional Neural Network

Accurate and efficient identification of vertebra fractures in spinal images is of utmost importance in improving clinical tasks such as diagnosis, surgical planning, and post-operative assessment. Previous methods that tackle the problem of vertebra fracture identification rely on quantitative morphometry methods. Standard six-point morphometry involves manual identification of the vertebral bodies’ corners and placement of points on identified corners. This task is time-consuming and requires effort from experts and technicians and prone to subjective errors in visual estimation in spinal images. In this paper, we propose an automated method to detect and classify vertebra fractures from 3D CT lumbar spine images. Fifteen 3D CT images with accompanying fracture labels for each of the five lumbar vertebra from the xVertSeg Challenge were utilized as data set. Each vertebra from the 3D image is processed into \(100\times 50\) 2D 3-channel images composed of three grayscale images. The three grayscale images represent the vertebral slices in the sagittal, coronal, and transverse anatomical planes. These \(100 \times 50\) 2D images are fed into the 152 layer Residual Network. A total of 13,400 images were generated from the data pre-processing stage. 12,700 of which having varying classifications were used as training data, and 100 images for each of the seven vertebra fracture classifications were used as testing data. The network achieved 93.29% testing accuracy.