Beam‐hardening correction in clinical x‐ray dark‐field chest radiography using deep‐learning‐based bone segmentation

Background Dark‐field radiography is a novel x‐ray imaging modality that provides complementary diagnostic information by visualising microstructural properties of lung tissue. Implemented via a Talbot–Lau interferometer integrated into a conventional x‐ray system, it permits simultaneous acquisition of perfectly registered attenuation and dark‐field radiographs. Clinical studies have shown that dark‐field radiography outperforms conventional radiography in diagnosing and staging pulmonary diseases, yet the polychromatic nature of medical x‐ray sources causes beam hardening and introduces structured artifacts, especially from ribs and clavicles. Purpose To address the artificial dark‐field signal arising from beam‐hardening and thereby improve the reliability of clinical dark‐field chest radiography by suppressing bone‐induced artifacts. Methods A segmentation‐based beam‐hardening correction (BHC) was developed that employs deep learning to segment ribs and clavicles and uses attenuation‐contribution masks derived from dual‐layer detector computed‐tomography data to refine the material distribution and estimate beam‐hardening effects. The rib segmentation network was trained on 196 chest radiographs with 49 validation images (VinDr‐RibCXR), and a clavicle network was trained on 56 images with 12 validation and 12 test cases. The trained models were applied to 174 dark‐field chest radiographs (51 chronic obstructive pulmonary disease, 86 COVID‐19, 37 healthy) and spectral CT scans from two patients; input data consisted of attenuation and dark‐field images and outputs were corrected dark‐field images and derived lung‐signal metrics. Results The proposed method markedly reduced bone‐induced artifacts and improved the homogeneity of the lung dark‐field signal. In comparative analyses, the corrected images exhibited diminished structured cross‐talk between attenuation and dark‐field channels, enhancing both visual interpretation and quantitative consistency across cohorts. Conclusions By combining deep‐learning‐based anatomical segmentation with material‐specific attenuation weighting, the proposed BHC suppresses the artificial dark‐field signal caused by polychromatic x‐ray spectra, leading to more reliable assessment of pulmonary microstructure in clinical dark‐field chest radiography.