Effects of different preprocessing factors on radiomics-based morphological brain networks

Radiomics has recently been applied to construct individualized morphological brain networks (IMBNs). However, how different preprocessing factors during the extraction of radiomic features affect the IMBNs remains largely unexplored. Using three independent datasets, here we compared the topological organization, test-retest (TRT) reliability, and disease susceptibility across IMBNs constructed based on radiomic features that were extracted from different image types (T1-weighted images versus gray matter volume maps) under three choices of preprocessing filters (no filter versus Laplacian of Gaussian filter versus wavelet filter). We found that regardless of the image types and preprocessing filters, the IMBNs exhibited non-trivial topological architecture (i.e., small-world organization and hubs) and fair-to-excellent TRT reliability. Nevertheless, gray matter volume-based IMBNs exhibited significantly higher TRT reliability than T1w-based IMBNs, and no filter- and Laplacian of Gaussian filter-based IMBNs showed significantly higher TRT reliability than Wavelet filter-based IMBNs. These results were robust when different brain parcellation schemes were employed for network node definition. In children with autism spectrum disorder, altered morphological connectivity and network topology were observed only in the IMBNs constructed from gray matter volume maps. Interestingly, the altered morphological connectivity was predominantly linked to the superior temporal gyrus and was significantly correlated with autistic traits, social skills, and behavioral difficulties of patients. Altogether, our findings indicate that different image types and preprocessing filters substantially influence the radiomics-based IMBNs in multiple aspects, thereby providing practical guidance for selecting appropriate strategies to obtain reliable results.