Application of radiomics-based image filtering to improve deformable image registration accuracy in thoracic images

Deformable image registration (DIR) is an important technique in radiation therapy. To improve DIR accuracy, we applied radiomics-based image filtering as a preprocessing step before DIR.

Thoracic four-dimensional computed tomography (CT) images of 10 patients with lung or esophageal cancer were examined. Peak-inhale and peak-exhale images were used for DIR. Before DIR, these images were converted into 90-voxel-based radiomics-based filtered images using extracted local radiomics features, respectively. On each filtered image, DIR between the peak-inhale and peak-exhale filtered images was performed. After DIR, the peak-inhale CT images were deformed to peak-exhale CT images using the displacement vector fields obtained from the DIR. The registration errors obtained from each radiomics-based DIR were calculated using landmark pairs and compared with conventional CT-based DIR using the same DIR parameters.

In radiomics-based DIR, the lowest registration errors (95th percentile) for intensity and texture features were 0.96 mm (right-left), 1.35-1.38 mm (anterior-posterior), 2.04-2.13 mm (superior-inferior), and 2.49-2.57 mm (three-dimensional). For CT-based DIR, the corresponding registration errors were 1.31 mm, 1.72 mm, 3.45 mm, and 3.98 mm.

By applying radiomics-based image filtering before DIR as a preprocessing, the registration error was lower than that of conventional CT-based DIR, suggesting that using radiomics may improve the accuracy of DIR.