The human brain structure changes with age throughout the lifetime (Ziegler et al., 2012). The atrophy of grey matter (GM) is commonly observed in normal brain aging, accompanied by the shrinkage of white matter volumes and enlargement of the cerebrospinal fluid spaces (Lemaitre et al., 2012). Trajectories of brain aging differ from individual, and aged brains are more prone to cognitive decline (Fjell et al., 2014). Thus, in the past few decades, researchers have continuously studied the macroscopic and microscopic manifestations of brain aging to varying degrees.

Voxel-based morphometry (VBM) and surface-based morphometry (SBM) are common approaches to studying brain morphological changes from the macroscopic level. VBM serves to estimate brain region volumes, such as grey matter volume (GMV) (Weise et al., 2019). SBM is applied to estimate a range of surface features, for example, cortical thickness (CTh), sulcal depth (SD), gyrification index (GI), as well as fractal dimension (FD) (Bachmann et al., 2023). These parameters can help distinguish between groups of controls and patients with neurological and psychiatric disorders (Nickel et al., 2019; Zhao et al., 2022; Ziukelis et al., 2022; Zhang et al., 2023). Previous cross-sectional and longitudinal studies have provided insights into brain region differences in morphological parameters of normal brain aging (Leong et al., 2017; Shen et al., 2018; Aljondi et al., 2019; Zheng et al., 2019; Lamballais et al., 2020). Research has also recommended that the combined use of VBM and SBM could better understand the brain neurobiological processes and improve the accuracy of morphological change detection (Goto et al., 2022).

Radiomics, as a rapidly developing field, can extract quantitative features from medical images to build diagnosis or prediction models to analyze microscopic information (Mayerhoefer et al., 2020). Radiomics features consist of region of interest (ROI) characteristics such as shape, first-order, and texture features, which can obtain a variety of unknown information from different modality images (Bang et al., 2021). The established disease-specific models could be potentially applied to solve clinical problems (Huang et al., 2023). In the area of neurodegenerative disorders, multiple radiomics models have been developed to diagnose mild cognitive impairment, Alzheimer’s disease, and Parkinson’s disease, and predict their progression and treatment effect (Bian et al., 2023; Shahidi et al., 2023).

Both VBM-SBM and radiomics are mainstream methods for neuroimaging analysis. The former focuses on comparing the macrostructural differences in brain regions to distinguish changes in diseases and the latter selects the most representative and meaningful features to build classification models. Although both methods are widely applied in neurodegenerative diseases, there is a lack of research comparing their performance in assessing brain aging. Therefore, this study aimed to construct a VBM-SBM model and a radiomics model based on the GM of cognitively normal (CN) individuals and compare the performance of the two models in identifying the degree of normal brain aging. By analyzing their strengths, weaknesses, and associations, our research could provide a reference for the selection and applicable situation of the two methods for future research.

VBM-SBM and radiomics methods are commonly used to analyze brain imaging from various perspectives. However, there has been no research comparing the two methods to explore their performance, associations, strengths, and applicable situations. In the current study, we built two normal brain aging models based on the GM of CN individuals, including a VBM-SBM model and a radiomics model. The radiomics model had significantly higher AUCs than the VBM-SBM model in the training, internal test, NACC test, and PPMI test sets. The correlations between age and four radiomics features were generally stronger than correlations between age and three VBM-SBM parameters, and weak relationships were found between the VBM-SBM parameters and radiomics features.

Previous researchers have used VBM and SBM to explore their relationships with age, however, this study applied the differential brain regions to the construction of a brain aging model. It was reported that the results of morphometric analysis varied with populations and thresholds, but there were still some frequently mentioned (Goto et al., 2022). In the literature on VBM research, it was found that the GMV of multiple brain regions significantly decreases with age, such as the temporal, occipital, and parietal lobes (Fleischman et al., 2014; Zheng et al., 2019). In our study, 5 regions from the temporal and parietal had significant differences between the middle-aged and old-aged groups. Finally, the right hippocampus, the largest cluster size, was selected to construct the model. In the SBM analysis, research on CTh showed that cortical thinning was more widespread than GMV loss (Fleischman et al., 2014). We found that the CTh of 12 regions had significant decreases with age, which could be observed in the temporal, occipital, parietal, and frontal lobes. This finding was consistent with previous research (Fjell et al., 2009). After the selection through logistic regression, the left lingual participated in the model construction. GI, a quantification of cortex folding structure, changes with aging and relates to the development of cognitive function across the lifespan (Cao et al., 2017). Studies have reported that global cortical gyrification gradually decreases with age, while significantly differential regions for local cortical gyrification vary in sample size and age distribution (Lamballais et al., 2020; Madan, 2021). We found the 8 regions of the old-aged group had a significantly lower GI than that of the middle-aged group, and the bilateral insular regions accounted for a large proportion. The left insula was finally included to the model. Higher GIs in 2 regions from the temporal lobe of the old-aged group were also observed in this study.

The ranges of AUCs of the VBM-SBM model and radiomics model were 0.640 ~ 0.746 and 0.737 ~ 0.838, respectively. In the training set, the radiomics model had an AUC of 0.778, which was significantly higher than the VBM-SBM model (AUC = 0.697). For the test sets, the VBM-SBM model had an AUC range of 0.701 ~ 0.746 in the three external test sets, but an AUC of 0.640 in the internal test set. The radiomics model had an AUC range of 0.737 ~ 0.838 in four test sets. These indicate that the radiomics model can better distinguish the age-related groups and have higher generalization; however, it suffers from a deficiency in biological interpretability, and the selected features have some challenges in clinical application. For the VBM-SBM model, it can locate important brain regions and the parameters are often used in clinical practice, but its performance is slightly weak, and the processing is relatively cumbersome and time-consuming. Overall, both approaches have their advantages and disadvantages, and investigators can choose based on their study objectives. For researchers prioritizing performance and generalization, radiomics offers a good choice, while those emphasizing interpretability and clinical practice may find VBM-SBM more suitable.

By observing the correlations between age and VBM-SBM parameters, and age and radiomics features, we noted that the correlations between age and the four radiomics features were generally stronger than the correlations between age and the three VBM-SBM parameters. This suggests that the radiomics features have closer relationships with age than the VBM-SBM parameters in brain aging analysis. We also compared the relationships between the VBM-SBM parameters and radiomics features, the results showed that though the indicators from the two methods had significant correlations, the associations were relatively weak, which means that VBM-SBM and radiomics analyze brain aging from different dimensions and perform their respective functions.

Several limitations in our investigation need to be acknowledged. Firstly, the retrospective data were from four databases, which had inherent biases in selecting subjects and might not represent real-world situations. Secondly, the age span between groups was relatively narrow, which could limit the transferability of the model performance. Thirdly, the selection thresholds of VBM-SBM parameters and radiomics features were based on the current study, which can be varied according to different research goals and situations. Finally, there are many methods and software available for morphometry and radiomics analysis of the brain. Therefore, the results of the model performance might change under different experimental conditions.

In conclusion, the radiomics model performed better than the VBM-SBM model. Radiomics focuses on the generalization and VBM-SBM has the interpretability. VBM-SBM and radiomics analyze brain aging from different dimensions and perform their respective functions.

Publicly available datasets were analyzed in this study. This data can be found at: the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu/), the Australian Imaging, Biomarker and Lifestyle (AIBL) database (https://aibl.org.au/), the National Alzheimer’s Coordinating Center (NACC) database (https://naccdata.org/), and the Parkinson’s Progression Markers Initiative (PPMI) database (https://www.ppmi-info.org/).

The studies involving humans were approved by the Ethics Committees of Alzheimer’s Disease Neuroimaging Initiative (ADNI), Australian Imaging, Biomarker and Lifestyle (AIBL), National Alzheimer’s Coordinating Center (NACC), and Parkinson’s Progression Markers Initiative (PPMI). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

YY: Conceptualization, Data curation, Formal analysis, Methodology, Writing – original draft, Writing – review & editing. XH: Conceptualization, Data curation, Methodology, Writing – review & editing. YX: Data curation, Formal analysis, Writing – review & editing. JP: Formal analysis, Visualization, Writing – review & editing. FZ: Formal analysis, Writing – review & editing. YS: Conceptualization, Data curation, Formal analysis, Funding acquisition, Methodology, Supervision, Writing – review & editing.