Shape-Aware Deep Convolutional Neural Network for Vertebrae Segmentation

Abstract

Shape is an important characteristic of an object, and a fundamental topic in computer vision. In image segmentation, shape has been widely used in segmentation methods, like the active shape model, to constrain a segmentation result to a class of learned shapes. However, to date, shape has been underutilized in deep segmentation networks. This paper addresses this gap by introducing a shape-aware term in the segmentation loss function. A deep convolutional network has been adapted in a novel cervical vertebrae segmentation framework and compared with traditional active shape model-based methods. The proposed framework has been trained on an augmented dataset of 26370 vertebrae and tested on 792 vertebrae collected from a total of 296 real-life emergency room lateral cervical X-ray images. The proposed framework achieved an average error of 1.11 pixels, signifying a 36% improvement over the traditional methods. The introduction of the novel shape-aware term in the loss function significantly improved the performance by further 12%, achieving an average error of only 0.99 pixel.