Comparative evaluation of automated MRI-based brain volumetry software for estimating cognitive domains in early Alzheimer’s disease

Abstract

Background

Brain volumetry software is widely accepted for assessment in Alzheimer’s disease. However, direct comparisons for software-specific prediction capabilities across cognitive domains have not been reported. This study evaluates the performance of four brain volumetry software programs in predicting domain-specific cognitive scores in patients with early Alzheimer’s disease (AD) using a consistent deep learning model.

Methods

A total of 255 patients with amyloid PET-confirmed AD were retrospectively enrolled. Brain volumetric features were extracted from 3D T1-weighted MRI using four software programs (AQUA, DeepBrain, A-finder, and FreeSurfer). A multi-layer perceptron model incorporating seven clinical variables and software-specific volumetric features was trained to predict six cognitive outcomes: the Mini-Mental State Examination (MMSE) score and five Seoul Neuropsychological Screening Battery (SNSB) domain scores. Model performance was evaluated using mean squared error and Pearson’s correlation coefficient (r).

Results

The FreeSurfer-based model showed the numerically highest correlation for MMSE (r = 0.56), language (r = 0.38), visuospatial (r = 0.21), and frontal/executive functions (r = 0.32). The AQUA-based model showed the numerically highest correlation for attention (r = 0.40), and DeepBrain for memory (r = 0.38). MMSE scores were generally better predicted than domain-specific scores across all models.

Conclusion

Brain volumetry software showed modest, domain-dependent associations with cognitive scores in early AD, with the strongest signal for MMSE and no statistically significant prediction for visuospatial function. These findings support the potential of brain volumetry software to partially estimate domain-level cognitive profiles, while differences across software should be interpreted with caution rather than as evidence of intrinsic software superiority.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02535-5.

Introduction

Brain volumetry refers to the automatic labeling, quantification, and visualization of brain structures from a set of MR images, enabling objective assessment of the brain, including atrophy [1]. Machine learning (ML)-based volumetry software such as FreeSurfer have been developed. After the rapid advancement in deep neural networks, multiple deep learning (DL)-based brain volumetric software have been developed and have received premarket notification from the U.S. Food and Drug Administration [2]. Using atlas-based automated segmentation and quantification of structural brain MR images, these software have the potential to aid clinicians in the diagnosis of neurodegenerative conditions, including Alzheimer’s disease (AD) [3]. These approvals are supported by many studies demonstrating the diagnostic and prognostic capabilities of brain volumetric methods for AD [4, 5]. Crane et al. showed that cognitively defined AD subgroups have distinct gray matter volume atrophy patterns, though their study noted limitations regarding ethnic diversity [6].

The Seoul Neuropsychological Screening Battery (SNSB) is widely used in Korea for assessing five key cognitive domains—attention, language, visuospatial, memory, and frontal/executive functions—in patients with neurological disorders [7]. Along with the MMSE, SNSB scores serve as critical indicators of cognitive function in AD, and their accurate estimation is essential for guiding clinical decisions. While brain MRI volumetry has shown promise in detecting structural changes associated with AD, its ability to predict cognitive function remains largely unvalidated. Moreover, given the known discrepancies between different volumetry software programs [8], it is unclear whether prediction performance differs depending on the software used. A direct comparison of software-specific prediction capabilities across cognitive domains is therefore warranted.

We hypothesized that cognitive domain scores in AD patients could be predicted from brain volumetric data using a consistent deep learning model, and that prediction performance would vary by software depending on differences in segmentation.

Materials and methods

Study participants

This retrospective study reviewed 1,855 patients suspected of having Alzheimer’s disease who visited the Dementia Outpatient Clinic of our institution between January 2014 and January 2022. Of these, 283 met all inclusion criteria: (1) brain Aβ deposition confirmed by ¹⁸F-florbetaben positron emission tomography (PET); (2) mild cognitive impairment (MCI) or early dementia with a Clinical Dementia Rating (CDR) of either 0.5 or 1; (3) completion of comprehensive neuropsychological testing, including the SNSB; and (4) availability of at least one brain MRI scan. Amyloid-positive individuals with CDR 0 were not included because this study was designed to model domain-level cognitive performance in clinically symptomatic early AD (CDR 0.5–1); cognitively intact amyloid-positive individuals represent the preclinical stage, in which SNSB domain scores largely fall within the normative range and provide limited variance for supervised regression of cognitive outcomes.

The exclusion criteria were as follows: (1) patients with other degenerative causes of dementia including frontotemporal dementia, dementia with Lewy bodies, corticobasal degeneration, and progressive supranuclear palsy (n = 3); (2) those with drug-induced cognitive impairment (n = 3); (3) the presence of other causes of cognitive impairment such as epilepsy, psychiatric disorders, normal-pressure hydrocephalus, or structural brain lesions (e.g., tumor or hemorrhage) (n = 7); (4) patients whose MRI scans lacked 3D T1-weighted imaging (n = 10); or (5) those whose MRI scans could not be analyzed by the software (n = 5). A total of 255 patients were included in the final analysis. Figure 1 summarizes the patient selection process and dataset division.

Fig. 1.

Flowchart of patient selection and dataset division; AD, Alzheimer’s disease; MMSE, Mini-Mental State Examination; SNSB, Seoul Neuropsychological Screening Battery; Val, validation

Clinical and neuropsychological assessment

Demographic and clinical data, including age, sex, education, and comorbidities such as hypertension, type 2 diabetes, dyslipidemia, and history of CVA (cerebrovascular accident), were collected through a review of patient medical records.

All participants completed the standardized SNSB [9]. Details of the test items are summarized in Supplementary Data 1. Standardized Z-scores of each test item were available based on age- and education-matched norms for all tests with scores. Z-scores were summed according to domain characteristics to yield scores for five domains: attention, language, visuospatial, memory, and frontal/executive function [9–12]. Each cognitive domain score was operationally defined as the sum of the available standardized z-scores assigned to that domain, with equal weighting of available tests within each domain. Binary frontal/executive items were not included. The full list of component tests used to construct the attention, language, visuospatial, memory, and frontal/executive domain scores is provided in Supplementary Table 1. Domain assignment followed the predefined analytic framework of SNSB used in this study. Although some component tests may engage overlapping cognitive demands (e.g., the Rey Complex Figure Test copy task involves executive and working memory components in addition to visuospatial processing), domain assignment was kept consistent with the predefined SNSB framework without further reassignment; this overlap is further addressed in the Discussion.

Image acquisition and processing

All participants underwent MR scans using a 3T unit (Ingenia CX) and a 20-channel head coil at a single institution. T1-weighted volume images were obtained using 3D magnetization-prepared rapid acquisition, with a gradient echo (MPRAGE) sequence with the following acquisition parameters: axial acquisition with a 224 × 256 matrix; 256 × 256 reconstructed matrix; 220 mm field of view; 1.0 × 1.0 × 1.0 mm³ voxels; TE (echo time), 4.6 ms; TR (repetition time), 9.6ms; flip angle, 8°; scan duration = 209 s; and no slice gap. The T1-weighted images were processed using four automated volumetry software tools for parcellation and brain volume quantification. The four brain volumetric software packages used in this study were AQUA (Neurophet Ltd.), DeepBrain version 1.1.1 (VUNO Corporation), A-finder version 1.0.0 (Heuron Co. Ltd.), and FreeSurfer version 7.1.1 (Harvard, https://surfer.nmr.mgh.harvard.edu/fswiki/CorticalParcellation). Trial licenses were obtained for the three commercial software packages, and academic license was used for FreeSurfer. Processing times were: AQUA ~ 180 s, DeepBrain ~ 130 s, A-finder ~ 180 s per subject, and FreeSurfer ~ 7 h. The parcellated volumes were normalized to the total intracranial volume.

Development of prediction models

A DL-based multi-layer perceptron (MLP) architecture was implemented to efficiently handle the high dimensionality of brain volumetric data with more than a hundred input variables and to obtain multivariable outputs from a single trained model. The MLP model consisted of two hidden layers. The first hidden layer comprised 60 nodes and the second hidden layer comprised 30 nodes, with dropout regularization (dropout rate = 0.2) applied to prevent overfitting. Mean squared error (MSE) was used as its loss function, and stochastic gradient descent (learning rate of 2e-3) served as the optimizer. The batch size for training was set to 256 and early stopping was implemented to avoid overfitting and ensure optimal performance. The development environment included Python 3.9, TensorFlow 2.10, and CUDA 11.2. The same MLP architecture and hyperparameters (number of hidden units, dropout rate, learning rate, batch size) were applied to all four software packages to enable a direct, controlled comparison across software packages, although the input dimension naturally varied across models (112–179 volumetric features depending on the software, as specified above). Software-specific hyperparameter optimization was not performed. Figure 2 illustrates the MLP-based model architecture for predicting six cognitive function scores. The code used for the model development is available at https://github.com/siksimi/Volumetry_MLP/blob/main/main.py.

Fig. 2.

Multi-layer perceptron architecture for multi-output regression predicting six cognitive function scores; CVA, cerebrovascular accident; DM, diabetes mellitus; HTN, hypertension; MMSE, Mini-Mental State Examination

The dataset, comprising 255 participants, was divided into training (n = 178, 70%), validation (n = 26, 10%), and test (n = 51, 20%) sets. This division was conducted temporally on August 20, 2018, which served as a natural cutoff point to yield an approximate 8:2 ratio between earlier and later cases. This approach reflects real-world clinical practice, where predictive models are typically applied to future patients, and also prevents information leakage by ensuring that the test set only includes patients enrolled after model development began. After dataset splitting, the training data was replicated 20-fold, and element-wise multiplicative Gaussian noise sampled from N(1, 0.04²) was applied to the replicated training data before scaling. This procedure was intended as a regularization strategy rather than to increase the effective sample size. The multiplicative formulation was used to preserve the proportional relationship among features with markedly different numerical ranges, and the fold size and noise magnitude were empirically chosen to mitigate overfitting of the multi-output MLP on a limited training set without introducing excessive perturbation. The data were then standardized using a standard scaler. Scikit-learn (version 0.21.2) and imbalanced-learn (version 0.9.1) libraries were used for dataset division and data standardization.

Four prediction models were trained based on the results of each brain volumetry software. The model inputs consisted of features extracted from brain volumetric software outputs and clinical variables. Specifically, the number of volumetric features included are as follows: 156 from AQUA, 179 from DeepBrain, 169 from A-finder, and 112 from FreeSurfer. Seven clinical variables were also included: age, sex, education, history of hypertension, type 2 diabetes, dyslipidemia, and history of CVA, given that cognitive function is influenced not only by structural brain atrophy but also by demographic characteristics and comorbidities. Specifically, age and education were incorporated because they remain associated with SNSB domain performance in early AD despite the use of age- and education-adjusted norms, while vascular risk factors (hypertension, diabetes, dyslipidemia, and history of CVA) were included as established contributors to cerebrovascular burden and mixed-pathology cognitive decline that may not be fully captured by volumetric atrophy alone. The model outputs were six cognitive scores, including the MMSE score and five domains of the SNSB, as detailed above (attention, language, visuospatial function, memory, and frontal/executive function). A clinical-only model, using only the seven clinical variables as inputs, was additionally trained as an ablation baseline to quantify the incremental predictive contribution of the volumetric features.

Statistical analysis

The clinical information of the patients was compared using the t-test, Chi-squared test, or Fisher’s exact test, depending on the type of information analyzed. The performance metrics of the prediction models were evaluated by analyzing the correlation between the predicted and actual cognitive scores using the Pearson correlation coefficient (r) and the MSE with 95% confidence intervals. Shapley additive explanations (SHAP) [13] analysis was conducted to estimate the contribution of each clinical and volumetric feature of the top-performing models for each cognitive score. Statistical analyses were performed using the R software (version 3.5.1; R Foundation for Statistical Computing), with statistical significance set at P < 0.05.

Results

Patient characteristics

A total of 255 patients were included in this study. The median age [1st quartile, 3rd quartile] was 75.1 years [70.1, 79.0] and the median years of education was 12 years [6.0, 14.0]. A total of 160 patients (62.8%) were female. Hypertension, diabetes mellitus, and dyslipidemia were present in 142 (55.7%), 54 (21.2%), and 62 (24.3%) patients, respectively. Furthermore, 18 patients (7.1%) had a history of cerebrovascular accidents. The median MMSE score was 24 [22, 26]. Overall, 218 patients (85.5%) had a CDR score of 0.5 and the rest had a CDR score of 1. The median sum of Z-scores for individual cognitive domains were as follows: attention score was − 0.5 [-1.6, 0.6], language score was − 0.7 [-1.7, 0.0], and visuospatial function score was − 1.3 [-3.1, 0.0]. The mean ± standard deviation of the sum of Z-scores for memory was − 7.7 ± 4.9, while the median sum of Z-scores for frontal/executive functioning was − 3.4 [-5.2, -1.5]. Table 1 summarizes the demographic and cognitive characteristics of patients in the training, validation, and test sets.

Table 1.

Demographic and cognitive characteristics in the training, validation, and test set

Variables	Training set (n = 178)	Validation set (n = 26)	Test set (n = 51)	P values
Clinical features
Age*	75.5 [70.7, 79.0]	73.8 [68.3, 78.8]	74.7 [69.8, 78.9]	0.649
Sex				0.378
Female	107 (60.1)	17 (65.4)	36 (70.6)
Male	71 (39.9)	9 (34.6)	15 (29.4)
Hypertension	104 (58.4)	13 (50.0)	25 (49.0)	0.406
Diabetes	34 (19.1)	7 (26.9)	13 (25.5)	0.463
Dyslipidemia	39 (21.9)	9 (34.6)	14 (27.5)	0.312
History of CVA	11 (6.2)	2 (7.7)	5 (9.8)	0.625
Education (years)*	12 [6, 15.8]	11 [6.8, 12]	9 [6, 12]	0.618
Cognitive functions
CDR				0.176
0.5	155 (87.1)	19 (73.1)	44 (86.3)
1	23 (12.9)	7 (26.9)	7 (13.7)
MMSE*	24.5 [22, 26]	24 [23, 25.8]	23 [21, 25]	0.189
SNSB
Attention*	-0.47 [-1.54, 0.72]	-0.58 [-1.72, 0.84]	-0.88 [-1.79, 0.28]	0.286
Language*	-0.68 [-1.61, -0.03]	-0.84 [-1.78, 0.16]	-1.08 [-2.10, 0.30]	0.486
Visuospatial*	-1.27 [-3.11, 0.00]	-1.57 [-3.50, 0.21]	-1.23 [-3.03, 0.03]	0.968
Memory†	-7.42 ± 5.17	-8.23 ± 4.58	-8.24 ± 4.01	0.425
Frontal/executive*	-3.33 [-5.04, -1.40]	-3.19 [-5.91, -1.36]	-3.66 [-5.28, -2.01]	0.692

Unless otherwise specified, data are numbers of patients, with percentages in parentheses

^*Data are median [1st Quartile, 3rd Quartile]

^†Data are means ± standard deviations

CVA, cerebrovascular accident; CDR, Clinical Dementia Rating; MMSE, mini-mental state examination; SNSB, Seoul Neuropsychological Screening Battery

Performance of brain volumetry-based prediction models

The performance of prediction models based on the four brain volumetry software packages (AQUA, DeepBrain, A-finder, and FreeSurfer) in predicting cognitive function scores is summarized in Table 2.

Table 2.

Performance metrics of brain volumetry software for predicting cognitive functions

Cognitive Function	Metric	AQUA	DeepBrain	A-finder	FreeSurfer
MMSE	MSE	15.91	16.45	19.10	9.82
	Pearson’s R P-value	0.397(0.136−0.606) 0.004	0.205 (-0.074−0.455) 0.148	0.282 (0.007−0.517) 0.045	0.559 (0.335−0.723) < 0.001
Attention	MSE	2.35	3.02	3.35	2.92
	Pearson’s R P-value	0.404 (0.145−0.612) 0.003	0.064 (-0.216−0.333) 0.658	-0.085 (-0.352−0.196) 0.556	0.004 (-0.272−0.279) 0.978
Language	MSE	3.50	3.35	3.88	2.85
	Pearson’s R P-value	-0.057 (-0.328−0.222) 0.689	0.056 (-0.224−0.326) 0.699	-0.224 (-0.471−0.055) 0.113	0.375 (0.111−0.590) 0.007
Visuospatial	MSE	7.81	7.70	8.34	7.50
	Pearson’s R P-value	0.093 (-0.188−0.359) 0.519	0.161 (-0.120−0.418) 0.261	0.038 (-0.240−0.310) 0.792	0.210 (-0.070−0.459) 0.139
Memory	MSE	16.03	14.28	18.16	15.72
	Pearson’s R P-value	0.254 (-0.024−0.495) 0.073	0.380 (0.116−0.593) 0.006	0.176 (-0.104−0.431) 0.216	0.311 (0.039−0.540) 0.027
Frontal/executive	MSE	7.36	8.17	7.83	7.03
	Pearson’s R P-value	0.242 (-0.036−0.485) 0.088	0.122 (-0.159−0.384) 0.395	0.133 (-0.148−0.394) 0.352	0.319 (0.048−0.547) 0.022

The confidence interval is indicated in parentheses for Pearson’s R

MMSE, mini-mental state examination; MSE, mean squared error

For the MMSE, frontal/executive, and language scores, the FreeSurfer-based model showed the lowest MSEs (9.82, 7.03, and 2.85, respectively) and the numerically highest Pearson’s r values (0.559, 0.319, and 0.375, respectively), with significance (P < 0.001, P = 0.022, and P = 0.007, respectively). For the attention domain, the AQUA-based model achieved the highest correlation (MSE = 2.35; r = 0.404; 95% CI: 0.145–0.612; P = 0.003). For the memory, the DeepBrain yielded the highest correlation (MSE = 14.28; r = 0.380; 95% CI: 0.116–0.593; P = 0.006). For the visuospatial domain, no model achieved a significant correlation between predicted and actual scores. The highest r was with FreeSurfer (r = 0.210; P = 0.139). Average Pearson’s r across all models was highest for MMSE (mean Pearson’s r: 0.360), followed by memory (0.280), frontal/executive (0.203), visuospatial (0.125), attention (0.096), and language (0.040). The A-finder-based model showed relatively lower correlations across domains. Figure 3 shows scatterplots of the correlations of the FreeSurfer-based model predicting the six cognitive function scores in the test set.

Fig. 3.

Linear correlations of FreeSurfer-based model predictions for MMSE and five SNSB cognitive domains; MMSE, mini-mental state examination; SNSB, Seoul Neuropsychologic Screening Battery

Supplementary Figs. 1 and 2 show scatterplots of the correlations between the models on the training and test sets, respectively. Figure 4 shows a heatmap of Pearson’s R between the predicted and actual cognitive function scores evaluated in the test set.

Fig. 4.

Correlation heatmap (Pearson’s r) of predicted and actual cognitive scores across four volumetry-based models and the clinical-only baseline; MMSE, Mini-Mental State Examination

The clinical-only model did not show statistically significant correlations for any cognitive outcome and yielded Pearson’s r values of 0.18 for MMSE, 0.15 for attention, 0.03 for language, 0.21 for visuospatial function, 0.17 for memory, and 0.19 for frontal/executive function (Fig. 4). Relative to this baseline, the comparison was domain-dependent: for MMSE and memory, all four volumetry-based models yielded numerically higher correlations than the clinical-only model, whereas for attention, language, and frontal/executive function only one or two volumetry models exceeded the clinical-only baseline and the remaining models performed at or below it. For the visuospatial domain, no volumetry-based model exceeded the clinical-only baseline. Formal statistical comparison between correlation coefficients was not performed.

The SHAP analysis results obtained from the top-performing models for predicting each cognitive score (except visuospatial function) are summarized in Supplementary Table 2. In the attention domain, the AQUA-based model identified key frontal and subcortical structures, including the left pallidum (mean SHAP = 0.081), left amygdala (0.080), and left caudal middle frontal cortex (0.069). The DeepBrain-based model for memory highlighted right temporal lobe structures, such as the middle temporal gyrus (0.24), right fusiform gyrus (0.22), and total temporal lobe volume (0.22). For language prediction, the FreeSurfer-based model emphasized classical language-related regions, including the left pars orbitalis cortex (0.089) and right superior temporal cortex (0.074). In the frontal domain, the FreeSurfer-based model highlighted the left insula cortex (0.15), left lateral orbitofrontal cortex (0.12), and right posterior cingulate cortex (0.11). For MMSE, the FreeSurfer-based model identified the right fusiform cortex (0.60), left inferior temporal cortex (0.57), and left pericalcarine cortex (0.57) as the most influential features, along with education (0.61) and age (0.33) among clinical variables.

Representative cases are shown in Fig. 5 and Supplementary Fig. 3. Figure 5 presents an example of MMSE prediction using the FreeSurfer-based model in a patient in his seventies (actual: 17; predicted: 16.8), with the right fusiform and left inferior temporal cortices contributing most strongly.

Fig. 5.

Illustrative case of MMSE score prediction based on volumetric brain MRI features using FreeSurfer; A patient in his seventies with 3 years of education and a Mini-Mental State Examination (MMSE) score of 17 was evaluated using the FreeSurfer prediction model, which estimated a score of 16.8. (A) Automated segmentation output highlighting brain structures extracted using FreeSurfer. (B) Top five regional brain volumes contributing to the predicted MMSE score, visualized on 3D cortical surfaces. Brain regions are color-coded and ranked by mean SHAP (SHapley Additive exPlanations) values, which indicate each region’s relative importance in the model’s prediction: right fusiform cortex (mean SHAP = 0.60), left inferior temporal cortex (mean SHAP = 0.57), left pericalcarine cortex (mean SHAP = 0.57), left insula cortex (mean SHAP = 0.52), and right superior parietal cortex (mean SHAP = 0.50)

Supplementary Fig. 3 illustrates attention prediction using the AQUA-based model, highlighting frontal structures consistent with established attention networks.

Discussions

We found that integrating brain volumetric data and clinical features showed modest associations with the cognitive scores of patients with AD across five domains (memory, visuospatial function, attention, frontal/executive, and language function), as well as the MMSE score, with Pearson’s r ranging from 0.21 to 0.56. Overall, the models performed relatively well in estimating MMSE and memory domain scores, while language function was more difficult to predict.

Prediction performance varied across volumetry software, with different programs showing distinct strengths in specific domains. The FreeSurfer-based model showed the numerically highest correlation for four out of six cognitive scores: MMSE, language, visuospatial, and frontal/executive function scores. By contrast, the AQUA-based model showed the strongest association with attention, and the DeepBrain-based model with memory. The numerically highest overall accuracy was observed for MMSE prediction using the FreeSurfer-based model, although pairwise comparisons between models were not formally tested. However, this study does not suggest that any specific volumetry software is intrinsically superior for assessing particular cognitive domains. Rather, by applying neural network architecture to outputs from different software, we aimed to examine how technical variability in segmentation influences prediction performance. This indirect comparison provides insights into how structural measurement differences translate into functional predictions, offering a practical framework for evaluating AI-based volumetric tools from a clinically relevant perspective. Importantly, the observed performance differences across software packages should not be interpreted as evidence of intrinsic superiority of any specific package. Rather, they may reflect differences in segmentation/parcellation implementation, anatomical definitions, feature granularity, hemisphere handling, and normalization conventions across software, as well as the way these software-specific volumetric outputs interact with the fixed MLP architecture employed in this study. Most brain volumetry software have been developed to assist in the diagnosis of AD by quantifying regional brain atrophy. However, domain-specific cognitive evaluation is essential in clinical practice. Extending the role of volumetry tools to provide indirect information on cognitive domains may enhance their clinical utility. Radiologists could play a key role in guiding such application-driven advancements.

Previous studies have attempted cognitive function prediction using neuroimaging data. With fMRI, Kwak et al. reported correlation coefficients ranging from 0.08 to 0.44 when using connectivity models with clinical variables [14], while another study using autoencoder connectivity models showed promise for predicting various cognitive domain scores [15]. Using structural MRI, Hojjati et al. successfully employed region volumes (middle temporal gyrus, hippocampus, entorhinal cortex, ventricles, fusiform gyrus) as features for neural networks predicting neuropsychological scores in AD [16]- with outcomes comparable to our study, suggesting potential clinical utility.

The gap between predicted and actual cognitive functions may be partly attributed to factors not captured in our model. One such factor is cognitive reserve – the brain’s resilience to neuropathological damage – which can vary widely between individuals. While our model incorporated years of education as a proxy, other determinants such as APOE4 genotype, subclinical neuropathological burden, and occupational or social engagement were not included [17, 18]. Furthermore, although comorbid pathologies such as vascular changes or Lewy body disease are prevalent among AD patients and may influence cognitive outcomes [19, 20], their confounding effects were likely minimized in our study by enrolling only patients with amyloid PET-confirmed AD and by excluding other neurodegenerative or secondary causes of dementia.

Despite the overall moderate performance, the prediction of visuospatial function remained challenging. None of our models achieved a statistically significant correlation with visuospatial function scores. This can be attributed to several factors. First, visuospatial processing involves widely distributed networks across occipital, parietal, and frontal regions with considerable individual variability in activation patterns [21, 22]. Second, neuroanatomical studies have shown that visual processing areas are relatively spared in early AD; Dickerson et al. reported AD-related cortical thinning occurs globally except in primary visual cortex [23], and Thiyagesh et al. noted functional deficits without corresponding structural atrophy in dorsal visual pathways [24]. Third, the Rey Complex Figure Test used in SNSB places demands on multiple cognitive domains including executive function and working memory beyond pure visuospatial ability [25]. Finally, dorsal and ventral visual streams are differentially affected in AD, with the ventral stream often compensating for dorsal stream deficits [24, 26, 27] - this selective impairment pattern may be obscured when measuring visuospatial function as a single aggregated score in our analysis.

SHAP analyses confirmed that the models largely identified neuroanatomical correlates relevant to each cognitive domain. Memory prediction through DeepBrain appropriately prioritized temporal lobe structures, consistent with the established role of the temporal lobe in memory function [28]. Language prediction with FreeSurfer emphasized classical language-processing areas (pars orbitalis, Wernicke’s area), which are recognized components of language-eloquent area [29]. For attention, the AQUA-based model highlighted key frontal and frontoparietal regions such as the caudal middle frontal cortex and superior parietal lobule, as well as subcortical regions like the pallidum and amygdala, reflecting distributed cortico-subcortical networks underlying human attention [30]. For the frontal/executive domain, the FreeSurfer-based model emphasized the insula, lateral orbitofrontal, and posterior cingulate cortices, regions implicated in salience detection, decision-making, and internally oriented cognition that are consistent with executive-control networks. For MMSE, influential features included the fusiform, inferior temporal, and pericalcarine cortices, reflecting the contribution of temporo-occipital recognition- and language-related regions to global cognitive performance in early AD. Despite variability in performance across domains, these results suggest that each model captured domain-relevant anatomical correlates, supporting the biological plausibility of the volumetric features used for prediction.

Despite its implications, our study has some limitations. First, we did not consider the segmentation processing times of the volumetric software. FreeSurfer-based models generally demonstrated higher predictive performance; however, they require significantly longer processing times than DL-based software (7 h vs. a few minutes) [31]. Second, distinguishing the laterality of the volumetric data increased both the number of variables and computational burden. However, we deemed it more appropriate to include laterality information to account for the asymmetrical nature of brain function. Third, the multidimensional input data were not pre-processed through a separate feature selection process. Instead, we relied on the embedded feature-selection function of the MLP model. Fourth, we did not compare our model with other architectures. Most importantly, the absence of external validation limits the generalizability and robustness of our findings. Given the single-center retrospective design, validation in independent external cohorts with heterogeneous MRI protocols and patient populations is a prerequisite for clinical application. In addition, we applied a fixed MLP architecture across all four software packages without software-specific hyperparameter tuning. Because the number and structure of volumetric features differ across packages, part of the observed performance variation may reflect differences in model-input compatibility under a common analytic framework rather than intrinsic differences in volumetric output quality. Although statistically significant, the observed correlation levels are not sufficient for direct clinical decision-making and should be interpreted as exploratory associations.

Conclusions

In conclusion, brain volumetry software showed modest, domain-dependent associations with cognitive function in patients with early AD — numerically strongest for MMSE and not reaching statistical significance for the visuospatial domain. These findings support the potential of brain volumetry software to partially estimate domain-level cognitive profiles, while differences across software should be interpreted with caution rather than as evidence of intrinsic software superiority.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s disease

CI

Confidence interval

CDR

Clinical dementia rating

CVA

Cerebrovascular accident

DL

Deep learning

DM

Diabetes mellitus

fMRI

Functional magnetic resonance imaging

HTN

Hypertension

MMSE

Mini-mental state examination

ML

Machine learning

MLP

Multi-layer perceptron

MRI

Magnetic resonance imaging

MPRAGE

Magnetization-prepared rapid acquisition gradient echo

MSE

Mean squared error

PET

Positron emission tomography

r

Pearson correlation coefficient

SNSB

Seoul neuropsychological screening battery

SHAP

Shapley additive explanations

Author contributions

BS conceived the study, acquired funding, administered the project, and supervised the work. CYL curated the data, conducted the investigation, and drafted the original manuscript. YS performed the formal analysis, developed the methodology, and created the software and visualizations. BSY and YGL provided resources and validated the results. All authors reviewed, edited the manuscript, and approved the final version.

Funding

This study was supported by Samsung Medical Center Grant (#SMO124041) and a research grant from Bracco Imaging Korea (IIS-KR-0045). This research was also supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Korea government (MSIT and Ministry of Education) (grant numbers: RS-2026-25468417 and RS-2023-00237618). The funding bodies had no role in the design of the study, collection, analysis, interpretation of data, or in writing the manuscript.

Data availability

Due to ethical restrictions, the datasets generated and/or analysed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethical approval

This study was conducted in accordance with the Declaration of Helsinki. It was approved by the Institutional Review Board of Samsung Medical Center (IRB 2024-08-055). The requirement for informed consent was waived as this study was based on a retrospective chart review.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.