The Influence of Post-Acquisition Image Processing on a Radiomic Signature Constructed from Planar Images

For radiomics to be accepted as a definite tool in medicine, the outputs must be robust, repeatable and reliable. Image processing alters the quality of the input data which might have an impact on the values of the extracted features and ultimately the signatures developed. This study evaluated the magnitude of the influence of various interpolation and post-acquisition processing methods on the radiomic feature values extracted from planar images and radiomic signatures. Three different interpolation methods were applied to a chest x-ray dataset before 2-dimesional (2D) radiomic features were extracted using Pyradiomics. The influence of image size, cropping and re-segmentation were also evaluated by changing the respective variable before applying bilinear interpolation and extracting 2D features. ANOVA and post-hoc Bonferroni corrections were used to assess the differences in the radiomic feature values. Of the 93 first order- and texture- features extracted, 42 texture features (56.8%) proved to be significantly influenced (p ≤ 0.05) by the interpolation method. Only 2 first order features (10.5%) were significantly influenced (p ≤ 0.05) by the image size and 62 texture features (83.8%) by the other pre-processing methods evaluated. Pearson’s Correlation Analysis was then applied to develop a separate radiomics signature from each of the six image processing datasets under consideration. Five identical signatures were developed, with only the uncropped dataset that resulted in a unique signature. This study showed that the interpolation algorithms and other processing applied to planar images do have a noticeable influence on most radiomic feature values extracted. But regardless of the differences seen in the feature values, the radiomic signatures were reproducible for most datasets using different image processing methods.