Deep Learning-Based CT Reconstruction Kernel Conversion in the Quantification of Interstitial Lung Disease: Effect on Reproducibility

Yura Ahn et al. analyzed the effect of CT reconstruction kernels on ILD quantification, using a deep learning-based kernel conversion algorithm to minimize variability. Involving 194 patients with ILD or interstitial lung abnormalities, the study utilized high-resolution CT images reconstructed with B30f, B50f, and B60f kernels, with B60f serving as the reference standard. Automated quantification was performed via Aview software, and reproducibility was evaluated using ICC and Bland-Altman methods. Results showed significant variability in disease pattern quantification across kernels, notably an overestimation of ground-glass opacity on smoother kernels. The kernel conversion algorithm effectively aligned measurements to the B60f reference, achieving nearly perfect ICC scores. This advancement enables consistent ILD assessments across different imaging protocols, enhancing reliability in multi-center studies. Limitations include the use of a single vendor’s scanner and unaddressed factors like slice thickness. Future research should focus on broader validation and inclusion of diverse CT parameters.