Author: Hyosung Cho, Dae Yup Han, Duhee Jeon, Jiwon Park, Hyesun Yang 👨🔬
Affiliation: Department of Therapeutic Radiology, Yale University School of Medicine, Yonsei University 🌍
Purpose: Scatter in X-ray imaging degrades image quality, hindering the visibility of critical anatomical features and complicating patient alignment in radiation therapy. This study aims to improve scatter correction using a modified DehazeNet architecture, originally designed for haze removal in computer vision. Enhanced kV-image quality can improve visibility of critical structures, enabling precise treatment planning and delivery with kV-triggered images.
Methods: The DehazeNet model was adapted to process kV X-ray images, refining transmission map and estimating atmospheric light for scatter correction. Key modifications included multi-scale convolutional layers for feature extraction and non-linear regression for scatter parameter refinement. Clinical evaluation used pelvic-prostate triggered imaging data, with parameters including a source-to-detector distance (SDD) of 1500 mm, source-to-object distance (SOD) of 1000 mm, X-ray tube voltage of 110 kVp, detector pixel size of 0.388 mm, and voxel size of 0.908 mm. Fiducial markers were analyzed for visibility and alignment precision. To evaluate DehazeNet performance on L-spine triggered images, pelvic phantom images were acquired at 3° gantry angle intervals using the Varian TrueBeam system. Additionally, clinical prostate-triggered images at 15° intervals were analyzed for the visibility of implanted gold seed markers. Contrast-to-noise ratio (CNR) was measured to assess image quality improvements.
Results: The modified DehazeNet demonstrated effective scatter correction across clinical scenarios. Fiducial markers obscured by scatter became clearly visible, enabling precise alignment with digitally reconstructed radiographs (DRR). Gold fiducial markers in the prostate region were accurately localized, improving imaging accuracy for treatment planning. The CNR for triggered images improved from 7.04 (scattered) to 8.78 (descattered, modified DehazeNet), confirming enhanced image quality.
Conclusion: This study highlights the potential of leveraging convolutional neural networks for scatter correction in X-ray imaging. By reducing scatter and enhancing critical imaging features, the method
facilitates accurate patient alignment and improves dose delivery in radiation therapy.