Stability of Features in Radiomics Studies with Application in Cancer Imaging

Radiomics allows additional information to be extracted from medical images, which cannot be seen with the naked eye. By calculating features based on the voxel intensities and spatial relationships, we can gain further knowledge of the underlying structure of cancer tumours. These radiomic features (RFs) can be utilised to predict cancer patient outcome. However, when patients are imaged there are a variety of settings and scanner manufacturers available, which can have a significant impact on RFs. Due to the findings in the literature there is much discussion on the reliability of RFs.

This work addresses two key radiomics questions. The first is whether consistent RF values can be obtained from a consistent experimental setup. The second is looking at CT image reconstruction parameters and how these effect RFs. A phantom, containing 10 different materials, designed for radiomic research, was CT scanned 5 times on the same scanner, with a 10-minute rest period between scans. The phantom was then scanned 10 times in quick succession, using the same setup. This tested the effects of the x-ray detectors heating up on the RF values. For the second research question, one CT image was reconstructed using a range of reconstruction parameters, such as kernel sharpness and denoise level, to assess their impact on RFs. RF stability was evaluated using a second order coefficient of variation.

X-ray detector heat was found to have a minimal impact on RF stability, with both sets of scans having 96% of RFs being stable. Across all the reconstruction parameters the RF stability was poor, with 69% of RFs being unstable. Reconstruction parameters, such as high kernel sharpness and denoising level were found to have a large impact on the RFs. These results help to determine which aspects of CT imaging need to be standardised to obtain reliable measurements in radiomic studies.