Effects of biplanar X-ray angular separation on deep learning-based volumetric CT reconstruction using X2CT-GAN

Abstract Accurate patient positioning is critical in image-guided radiotherapy (IGRT) to ensure precise dose delivery while minimizing radiation exposure to surrounding organs at risk. However, frequent or real-time volumetric imaging using cone-beam computed tomography (CBCT) is limited by additional dose, long acquisition time, and hardware constraints. To address these limitations, this study investigates a deep learning-based biplanar X-ray reconstruction framework designed to generate volumetric CT-equivalent images from only two X-ray projections. The proposed method employs a dual-view encoder-decoder architecture trained using paired digitally reconstructed radiographs (DRRs) and CT volumes synthesized from a public chest CT dataset. To determine the optimal biplanar configuration, projection angles ranging from 30° to 90° were systematically analyzed. Both qualitative and quantitative evaluations — including the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and contrast-to-noise ratio (CNR) — were performed to assess reconstruction accuracy and contrast preservation. The results demonstrate that the framework effectively reconstructs fine anatomical structures, including skeletal and fiducial features, from minimal projection data, with reconstruction quality improving as the angular separation increased. The best overall performance was achieved at an angular separation of 70°, yielding SSIM = 0.588, PSNR = 26.71 dB, MAE = 126.3, and CNR = 0.74, representing the optimal balance between complementary anatomical information and geometric fidelity. These findings suggest that an intermediate biplanar angle (∼ 70°) enables accurate and low-dose volumetric reconstruction, highlighting the potential of this approach for real-time, IGRT applications.