In this paper, we proposed a standard individualized brain structural similarity network based on Radiomics, and verified it with remarkable network/biomarker characteristics in test-retest reliability, graph theory attributes and individual identification, and finally exemplified it with potential performances in two representative applications: pMCI/sMCI discrimination and fluid intelligence prediction. The results implied that iRSSN provides a reasonable and unique manner to understand the structural organization of the brain, which could be extended into various diagnostic applications and individual trait predictions.
For the methodological considerations about iRSSN, several points should be underlined. First, VBM analysis is an essential and beneficial step for iRSSN, which is the core difference between Zhao’s method 16 and ours. The conventional Radiomics analysis is directly conducted on the whole brain image, which may contain some latent problems: 1. Inaccurate whole-brain registration into MNI template may induce bias in subsequent Radiomics calculation, while segmentation of gray matter could reduce the bias and thus improve the reliability; 2. the texture features in Radiomics may be influenced by the adjacent tissues such as white matter or skull. Taken together, VBM analysis can further improve the effectiveness of iRSSN, and the comparison between our method and Zhao’s method 16 in fluid intelligence also demonstrated that VBM analysis improved the performance and refined the selected feature sets. Second, radiomic features should be selected to construct iRSSN. If the number of features is too large, iRSSN will be densely connected since many of radiomic features are correlated each other. If only a few features are used, the reliability of iRSSN may be decreased. Third, smoothness is an essential step in common voxel-based morphometry and surface-based morphometry analysis, and it is also found to affect the test-retest reliability and the sMCI/pMCI classification in the study. Four, the feature normalization manner also influenced the sMCI/pMCI classification, implying the normalization manner should be carefully determined.
Although the straightforward physiological significance of iRSSN is unknown, radiomic features have been found to be linked with genetic, molecular, immune and histopathologic markers 17, indicating the possible intrinsic physiological significance of iRSSN. Moreover, iRSSN displays similar network characteristics as other well-known brain networks 39–41, such as the small-worldness, indicating it a type of efficient and economical brain network. With the increase of sparsity thresholds, the characteristic path length of iRSSN becomes lower while the clustering coefficient, local and global efficiency becomes higher, demonstrating iRSSN as a reasonable type of brain structural network. Moreover, the degree distribution of nodes in iRSSN shows an exponentially truncated power-law form, reflecting several hub regions exist in iRSSN. The hub regions mainly include precuneus, thalamus, basal ganglia, superior temporal gyrus, cingulate gyrus, middle temporal gyrus and so on, which are consistent with the hub findings 42,43 in other structural/functional networks of adult brain. These hubs are related to the specific structural module and involved in some key cognitive and behavioral domains, such as self-awareness 44, cognitive computation 45, speech processing 46, language 47 and prosocial behavior 48. Furthermore, the spatial distribution of iRSSN modules is largely different from the RSFC modules, and most modules of iRSSN only displayed a dice coefficient lower than 0.4 with RSFC. Therefore, iRSSN can provide a unique perspective to understand the brain organization complementary to RSFC, and the fluid intelligence prediction task also supported that combination of IRSSN and RSFC could largely improve the model performance, which is obviously superior to Zhao’s method 16. In future, further study can be made to investigate the structure-function coupling in each module of iRSSN in human development, cognition and ageing. Lastly, iRSSN displays typical characteristics of a good biomarker: high test-retest reliability and good fingerprinting performance, which are significantly better than RSFC, indicating its great potential in individualized clinical diagnosis.
To validate the potential applications of iRSSN, two representative tasks were selected including sMCI/pMCI discrimination and fluid intelligence prediction. To effectively recognize pMCI patients is clinically important but difficult, and most studies have just achieved classification accuracies lower than 0.80 32,33. The key point is that the features used for classification should sensitively reflect the tiny differences between sMCI and pMCI. In contrast, iRSSN obtained an acceptable accuracy of 0.80 and AUC of 0.859, and the discriminative features were mainly located at the ventral/dorsal attention network, default mode network and frontoparietal network. The networks were persistently reported abnormalities in MCI patients 34,49−51, and future study could combine iRSSN with other conventional brain networks to improve the diagnostic accuracy for pMCI patients. For fluid intelligence prediction, iRSSN displayed comparable accuracy with RSFC but used significantly fewer connectivity features, indicating iRSSN provides a simplified but effective representation for fluid intelligence. Specifically, combination of iRSSN and RSFC could promote the prediction performances, demonstrating iRSSN contains complementary information in comparison to RSFC, which offers a novel way to improve individual traits prediction. Taken together, iRSSN possesses a reliable, unique and sensitive perspective to descript the brain structural connectivity, and can be extended into various tasks of disease diagnosis and individual trait prediction.
Several limitations about the current study should be underlined. First, differences in the voxel size of scanned sMRI images exist between HCP and other clinical datasets, and it is unknown whether the original voxel size affects iRSSN. Second, the multi-center test-retest dataset in the study adopted more time-consuming MP2RAGE sequence compared to the conventional MPRAGE sequence, which may also affect iRSSN.