CONTENTS
1.1 Introduction: Background and Driving Forces
1.2 Materials and Methods
1.2.1 Dataset Descriptions
1.2.2 Isolation of the Spine
1.2.3 Detection of Vertebrae
1.2.4 Segmentation of Vertebrae
1.2.5 Measurement of the Cobb Angle
1.3 Results and Discussion
1.3.1 Evaluation of Segmentation
1.3.2 Evaluation of Cobb Angle Measurement
1.4 Conclusion and Future Work
References
1.1 INTRODUCTION: BACKGROUND AND DRIVING FORCES
The human spinal cord consists of 33 vertebrae named cervical (7), thoracic (12), lumbar (5), sacral (5) and coccygeal (4). Among them, the upper 24 vertebrae are movable, while the lower 9 are fixed [15]. Scoliosis refers to a condition where the otherwise straight human spine has a lateral curvature, forming an angle, and the spine becomes “C”- or “S”-shaped.
This angle, called the “Cobb angle” [14], is the angle formed between the most tilted vertebrae above and below the apex of the curvature and is the measure of the severity of the disease. A Cobb angle in the range of 10 to 20 degrees is deemed mild scoliosis; it is moderate when the angle is between 20 and 40 degrees and severe when the angle is more than 40 degrees [7]. The diagnosis of scoliosis is done by determining the Cobb angle through the analysis of computed tomography (CT) scans, X-ray scans or magnetic resonance imaging (MRI).
The need for surgery can be avoided with early diagnosis of the disease. However, the detection of scoliosis from X-ray or MRI images sometimes turn out to be extremely difficult due to morphological differences between patients. Also, there are limitations on the intensity of radioactive rays that can be used, which often results in underdeveloped radiograph images.
Deep learning is an essential component of artificial intelligence (AI), which gained popularity due to its ability to learn non-redundant, informative features on its own through back-propagation, unlike machine-learning techniques, where handcrafted features need to be extracted and selected manually [23, 12].
In the literature, there exist numerous computer-aided methods, but all of them have shortcomings. Mathematical filter-based segmentation methods like the active contour model [2] and charged-particle models [18] were proposed to segment the vertebrae before calculating the Cobb angle. These methods suffered from excess computational requirements and were susceptible to errors due to slight variations in images. Alharbi et al. [1] proposed a fuzzy spatial relation associated with deformable models to perform semi-supervised segmentation of the vertebrae. Recently convolutional neural network (CNN) and deep neural network models have been widely used for this purpose, with a significant improvement in the result [3]. Fu et al. [9] proposed a lightweight multitask network that primarily detects the corner points followed by accurate contour segmentation. In the literature, there exist other deep-learning based object-detection and segmentation methods [8, 6], but they suffer from the problem of extreme computation requirements.
In this chapter, we develop an automated method for the detection of vertebrae to measure the Cobb angle for scoliosis using deep learning and image processing. For this, first, the region of interest (RoI), that is, the vertebral column in this case, is extracted from the images using image processing. Then the vertebral column is segmented from the RoI by locating spine centers (SCs) and the spine boundary. Next, deep learning is used to segment the vertebrae from the vertebral column and, finally, the Cobb angle is measured using a method known as minimum bounding rectangle (MBR). In this research, we use the VGG-19 [19] convolutional neural network as the backbone for U-Net [17] architecture to perform semantic segmentation of the vertebral column from X-ray images. Semantic segmentation refers to the pixel-level classification of each image; that is, each pixel is classified as belonging to an object class or background class.
