Intervertebral Discs Localization and Segmentation Based on Broad Learning System and IPU-Net

Abstract

The intervertebral discs play an important role in jointing two vertebral bodies of the human spine. Accurate localization and segmentation of intervertebral discs are necessary for spine disease diagnosis, e.g., disc degeneration and scoliosis. However, manual methods are laborious and subject to subjective factors. In this work, we present an automatic method based on the broad learning system and a deep learning network for localization and segmentation of intervertebral discs from magnetic resonance images. Firstly, the broad learning system is trained to classify the patches of spinal images. Secondly, linear regression and fuzzy C-means are used to refine the localization result. Moreover, a proposed network named Inception Pooling U-network is used to segment the cropped intervertebral discs. The final segmentation result is obtained by combining the localization information and the intervertebral disc segmentation result. Quantitative comparisons are represented in terms of Dice coefficient with 0.9409 and Hausdorff distance with 13.1862, which is better than the result of compared methods. The experiment results significantly verify the effectiveness of the proposed method.