An Automated Cobb Angle Estimation Method Using Convolutional Neural Network with Area Limitation

Abstract

Cobb angle measurement is the gold standard for the idiopathic scoliosis assessment, and the measurement result is very important for the surgical planning and medical curing. Currently, the Cobb angle is measured manually by physicians. They find the four landmarks on each vertebra and calculate the Cobb angle by rules, which is time-consuming and unreliable. In this paper, we apply the convolutional neural network (CNN) to find the landmarks automatically based on anterior-posterior view X-rays, then output the Cobb angle results. The X-rays always have too much noise, which has a strong influence on the landmark estimation. Addressing this problem, we first detect each vertebra bounding box to provide an area limitation. Then we use the CNN with an enhancement module to find the landmarks on detected vertebra bounding boxes, which can remove the noise in the background. Our experiment results show that our two-stage framework achieves precise landmark location and small error on Cobb angle estimation. Therefore our method can provide reliable assistance for the physicians.