Learning Perspective Distortion Correction in Cone-Beam X-Ray Transmission Imaging

In cone-beam X-ray transmission imaging, perspective distortion (PD) causes difficulty in direct, accurate geometric assessments of anatomical structures. Since PD correction from a single view is highly ill-posed due to missing stereo/3-D information, the efficacy of different view combinations is investigated in this work. Our theoretical analysis reveals that the 0°&180° complementary view setting provides a practical way to identify perspectively deformed structures by assessing the deviation between the two views. In addition, it provides bounding information and reduces uncertainty for learning PD. Beyond view combinations, the impact of learning PD in different spatial domains, specifically Cartesian and polar coordinates, is explored. Two representative networks Pix2pixGAN and TransU-Net for correcting PD are investigated. Experiments on numerical bead phantom data and head CT data demonstrate the advantage of complementary views over other view combinations (a 0° single view, 0°&90° orthogonal views, and 0°&5° small angular views). Results further show that both Pix2pixGAN and TransU-Net achieve better performance in polar space than Cartesian space. The efficacy of the proposed framework on real cone-beam computed tomography (CBCT) projection data and its potential to handle bulky metal implants and surgical screws indicate the promising aspects of future real applications.