Introduction

The human cortical folding is highly convoluted in shape, which characterizes the inter-subject variability. Current studies found that the structural (e.g., BrainPrint¹) or functional (e.g., functional connectome fingerprinting^2,3) features of the adult brain are unique and reliable for individual identification. However, these studies were all performed on adult brains, which have relatively subtle changes across different scans. In contrast, the infant identification is more challenging, since the infant cortex develops dynamically and regionally-heterogeneously in the first postnatal years. Thus, the dedicated shape descriptor needs to be stable during the early brain development. To this end, we propose a novel morphological folding descriptor based on multi-scale curvatures, called FoldingPrint, for infant identification. This study thus aims to address two essential neuroscience questions: 1) whether the developing infant cortical folding is unique for individual identification; 2) which cortical regions are more distinct and reliable for infant identification.

Methods

Longitudinal T1w and T2w MR images of 472 healthy infants at 0, 1, and 2 years of age, with totally 1,141 scans, were acquired on a 3T Siemens scanner. All MR images were processed via an infant-dedicated computational pipeline⁴. Each reconstructed cortical surface was aligned onto the age-matched template in the UNC 4D infant cortical surface atlas⁵ to obtain a parcellation map⁶ with 34 regions of interest (ROIs). We created FoldingPrint through the following steps: first, the mean curvature map of each inner cortical surface was decomposed into multi-scales using over-complete spherical wavelets⁷, thus obtaining a comprehensive characterization of cortical folding. Second, in each ROI, the three most informative curvature scales (scales 2, 3, and 4 as shown in Fig. 1) were selected to construct a high-dimensional folding descriptor, i.e., FoldingPrint, based on their joint probability density as shown in Fig. 2. As we can see, the feature patterns on the 2D projections of the FoldingPrint are variable across different subjects, but stable across different ages of the same subject. Thus, the proposed descriptor can be used for characterizing the developing cortical folding for infant identification. Third, given a new scan to be identified, its FoldingPrint was compared to that of each scan in the existing database in each ROI. Herein, the Chi-squared distance was adopted to measure the similarity of each ROI’s FoldingPrint between different subjects. In each ROI, the subject with the highest similarity was regarded as the potentially-identified subject. In total, 68 results were obtained from all ROIs in both hemispheres. Finally, through an ROI-based majority voting scheme, the potentially-identified subject with the highest frequency was considered as the finally-identified subject. In addition, to explore which regions are more reliable for individual identification, we further computed the ROI-specific identification accuracy maps.

Results

The experiments were carried out in 3 tasks: 1) using the scans at birth to identify their corresponding scans at year 1; 2) using the scans at birth to identify their corresponding scans at year 2; 3) using the scans at year 1 to identify their corresponding scans at year 2. Since cortical folding develops more rapidly in the first year than the second year, the first two tasks are much more challenging than the third one. Table 1 reports the identification accuracies of different high-dimensional feature combinations. The results of the proposed FoldingPrint with multi-scale curvatures W₂-W₃-W₄ (10) are in bold. As we can see, compared to other different 2D combinations of decomposed curvatures, the proposed descriptor obtains the best performance on all three tasks, i.e., 98.97%, 98.58%, and 100.00%, respectively. Besides, from the results of high-dimensional feature combinations of the traditional cortical folding features, i.e., mean curvature, average convexity, and sulcal depth, we found that they still failed to show comparable performance as the proposed FoldingPrint, especially in the first two tasks. In addition, the maps of the region-specific accuracies in infant identification are displayed in Fig. 3. We found that the regions with high identification capability are mainly distributed in the high-order association cortices, especially in the prefrontal cortex.

Discussion and Conclusion

We proposed a novel morphological descriptor based on multi-scale decomposed curvatures, i.e., FoldingPrint, for infant identification. The proposed descriptor is capable of comprehensively characterizing the invariance of developing cortical folding and achieves promising identification accuracy. In addition, we found that the high-order association cortices, especially the prefrontal cortex, encode more individual variability, thus presenting high identification capability. Besides, the patterns of the ROIs’ identification capability are region-specific and time-consistent. Future work will include identifying the twins and exploring their individual variability patterns for better understanding the genetic and environmental influences on the individual variability of cortical folding.

Acknowledgements

This work was supported in part by NIH grants (MH100217, MH108914, MH107815, MH110274, MH070890, MH064065, HD053000, MH109773, MH116225 and MH117943).