Weighting Histogram of Oriented Gradients for Spondylolisthesis Classification from X-Ray Images

Abstract

Accurate spondylolisthesis classification is crucial for planning effective patient treatment. The machine learning technique is widely used to efficiently analyze spondylolisthesis from X-ray images. However, existing methods are sensitive to noise effects when using X-ray images. Moreover, the effectiveness of existing feature extraction is insufficiently accurate. Therefore, the weighting Canny histogram of oriented gradients (HOG) method is proposed to increase the accuracy of spondylolisthesis classification. This method uses an anisotropic filter to reduce noise in a pre-processing step. Then Canny operator is applied instead of x-and y-derivative filter of the HOG method to achieve better gradient images. After that, the slope value of the lumbar vertebra is calculated to weigh the texture HOG features. Thus, our features have properties of texture and shift of lumbar vertebrae. On the BUU Spine dataset, the weighting Canny HOG method yields high recall, precision, F1-score, and classification accuracy of 0.7488, 0.8526, 0.7832, and 0.9155. Our method is able to efficiently extract texture and shift features, thus improving the effectiveness of classifying spondylolisthesis from X-ray images.