AUTOMATIC SPINE AND PELVIS DETECTION IN FRONTAL X-RAYS USING DEEP NEURAL NETWORKS FOR PATCH DISPLACEMENT LEARNING

This paper proposes a method to automatically detect the spine and pelvis structures from a postero-anterior radiograph. From a training dataset, a non-linear regression model was trained using a deep neural network (DNN) in order to predict the displacement that recovers the optimal location of an anatomical landmark from an input image patch. Using a DNN for each landmark of a 2D simplified model of the spine, a detection sequence was able to localize the vertebral body centers and femoral heads. The whole process is regularized using a statistical shape model of a simplified model of the spine. The quantitative assessment on a set of 121 radiographs of scoliotic patients presented a mean localization errors of 3.5 +/- 3.6 mm and 5.7 +/- 6 mm respectively for the femoral heads and the vertebral body centers (vertebral levels T1 to L5). The mean error for the spinal curve automatic detection was 2 +/- 2.8 mm, which is accurate enough to determine a first estimate of the spine 3D reconstruction in a 3D biplanar reconstruction scheme.