Scatter Correction Using Deep Learning for Cone-Beam Computed Tomography

Scattering X-ray radiation is one of the most degradation factors in cone-beam computed tomography (CBCT) imaging, which has a significant impact on the contrast, sharpness, noise level, and background inhomogeneity of the image. X-ray photons that pass through an object are either absorbed or scattered, depending on the internal composition and structure of the object. A flat-panel detector is commonly applied in volumetric CBCT imaging which records both direct primary and scattered X-ray photons. Due to the large illuminated volume, a high proportion of scattered X-rays is measured in the total detected signal. The presence of scattered signals violates the backprojection assumption of a straight line integral and therefore is considered as artifact in the image formation. The image quality degradation due to scatter can be categorized as the following major folds: The first is the reduction of the image contrast within an object which may impair the visibility of low-contrast features. Second, scatter may degrade the accuracy of CT number for CBCT, causing cupping artifact in homogeneous regions and streaks within regions of high contrast. To correct for artifacts due to scatter contamination, scatter correction schemes are usually applied to subtract the scatter signal from total projection data. Noise in the corrected projection is magnified to further degrade the image quality due to the decreased signal magnitude. Extra noise suppression methods are thus required to suppress the magnified noise. These effects will reduce the CBCT as a quantitative imaging technique for many clinical applications. Scatter correction is therefore a critical aspect of advanced medical X-ray CBCT imaging. Though various scatter correction methods have been proposed in the literature, deep learning–based strategies attract the focus of researchers due to its feature extraction and representation capability over conventional non-learning-based schemes. In this chapter, we will review the state-of-the-art deep learning–based scatter correction methods and perform a deep analysis on their advantages in clinical applications.