According to the World Health Organization, the global population aged over 60 is projected to reach 2.1 billion by 2050, while the number of individuals over 80 is expected to triple to 426 million between 2020 and 2050 (1). Dementia is a prevalent condition, with its global incidence steadily increasing in recent years (2, 3). Timely diagnosis of conditions such as Alzheimer’s disease (AD) and mild cognitive impairment (MCI) is crucial for maintaining patients’ quality of life, as early intervention can potentially slow disease progression (4–6). Hippocampal atrophy is a characteristic feature of AD, making the recognition of this atrophy at an early stage particularly useful (7).

One commonly used image analysis tool is VSRAD (Voxel-based Specific Regional Analysis System for Alzheimer’s Disease), developed by Eisai, a computer-aided diagnostic system designed to support theclinical diagnosis of AD at an early stage (8, 9). VSRAD uses modified voxel-based morphometry (VBM) software. This software combines Statistical Parametric Mapping 8 (SPM8) and Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) algorithm for the detection and quantitative assessment of AD to compare a patient’s brain MRI to a normative database of healthy individuals (10). Thus, VSRAD primarily evaluates atrophy in the medial temporal lobe (including the hippocampus and parahippocampal gyrus), visualizing results as Z-score maps and quantifying the degree of atrophy in various brain regions with high versatility (11, 12). VSRAD is used in over 3,000 facilities in Japan and has been validated across multiple centers. This corresponds to roughly 30–40% of all MRI-equipped facilities nationwide. VSRAD ordinarily requires three-dimensional T1-weighted volume images, which take about five minutes to acquire. However, acquiring such data is often challenging in patients with suspected dementia due to poor compliance and motion artifacts, which frequently degrade image quality. Consequently, routine clinical imaging still relies mainly on two-dimensional sequences. However, the relatively large slice thickness of typical 2D scans (approximately 4–5 mm) poses a particular challenge for quantitative analyses like VBM, especially when analyzing historical 2D image data.

SynthSR is an image super-resolution technique recently implemented in FreeSurfer, a comprehensive open-source software package used for processing and analyzing brain MRI images. It has been shown to be capable of generating high-quality, high-resolution images from low-resolution MRI scans, which ultimately improves the accuracy of brain structure analysis (13, 14). Compared to conventional interpolation methods, SynthSR has been reported to reproduce more detailed brain structures and can complement older, low-resolution MRI datasets (15). Furthermore, SynthSR has served as a benchmark for newer models such as LoHiResGAN, which convert low-field into high-field equivalents (16), and has been integrated into workflows for Alzheimer’s disease assessment using hippocampal and white matter hyperintensity quantification (17). Recent studies have also applied SynthSR to ensure anatomical consistency in youth cohorts (18) and to generate cerebrospinal fluid (CSF) volumetrics predictive of stroke outcomes (19). These applications highlight SynthSR’s potential in both standardizing heterogeneous datasets and expanding the utility of legacy MRI data or data from resource-limited settings.

In this study, we evaluated the agreement and comparability of VSRAD analysis results between 3D volumes generated from 2D T1-weighted images and standard 3D T1-weighted volumes, with particular focus on between-method agreement, rank-based consistency, and diagnostic performance. We assessed whether SynthSR-generated 3D images from 2D inputs could yield results comparable to standard 3D images, potentially offering a more practical approach for dementia assessment.

Seventy-five patients underwent MRI during the study period. Of these, 21 (28.0%) were male. The mean age was 83.5 years (range, 61–107 years). This cohort reflects the typical population undergoing dementia screening in Japan. For all cases, 3D volume data (T1w_2DSR) were successfully generated from T1w_2D using SynthSR. The generated T1w_2DSR did not reconstruct the parietal CSF region because it was not imaged in the original T1w_2D. Each output image was independently reviewed by two neurosurgeons. Cases with obvious motion artifacts and disrupted WM and GM segmentation in the control T1w_3D images were excluded from the analysis. Among the excluded cases, four (Cases 22, 26, 43 and 63) had showed segmentation errors involving the ventricles and GM due to brain atrophy (Figure 3), five (Cases 7, 11, 12, 69 and 72) showed segmentation failure caused by motion artifacts; and two (Cases 15 and 39) had structural brain lesions due to stroke (Figure 4). A total of 64 cases were analyzed after excluding 11 cases. After VSRAD analysis, four scores and three volumes were calculated. For these evaluated indices, normality was not met in most cases (Shapiro–Wilk test, p < 0.05), and thus non-parametric comparisons were adopted.

Wilcoxon signed-rank tests were conducted across seven indices. As shown in Table 1, all comparisons showed statistically significant differences between the two methods. The proposed method tended to yield larger volume measurements in most VSRAD scores and segmented volumes. After applying Bonferroni correction for multiple comparisons (n = 7), all p-values remained below the significance threshold of 0.05, confirming the robustness of the findings.

Robust Bland–Altman analysis was conducted across seven indices. For each measurement pair, the median of the difference (Proposed method – Standard method) and the IQR were used to estimate robust limits of agreement defined as median ± 1.5 × IQR (Table 2). Among all indices, IQRs and derived limits of agreement varied across metrics, reflecting higher variability in volume-based measures than in Z-scores. Outlier analysis revealed that fewer than 10% of cases fell outside the robust limits of agreement for all variables, with the highest outlier proportions observed in gray matter volume and Score 1, both at 9.4%. Notably, score 3 “Ratio” and WM showed the largest absolute median differences, suggesting consistent deviations between the two methods (Figure 5).

Spearman’s correlation coefficients between the standard and proposed methods across all seven indices ranged from 0.40 to 0.74, with all values indicating statistically significant positive correlations (p < 0.001). Strong correlations (ρ > 0.7) were observed in score 3 “Ratio” and brain volume, while the remaining indices showed moderate correlation (0.3 < ρ < 0.7), indicating consistent ranking between the two methods (Figure 6).

Twenty-nine patients were diagnosed with AD, 8 were considered cognitively normal, and the rest were diagnosed with other dementias. ROC curves were drawn for AD and cognitively normal controls based on Score 1 “Severity” in VSRAD analysis. For the standard method (T1w_3D), AUC = 0.90, Youden Index = 0.72, and for the proposed method (T1w_2DSR), AUC = 0.96, Youden Index = 0.84 (Figure 7). DeLong’s test revealed no statistically significant difference between the AUCs of the standard and proposed methods (Z = 0.009, p = 0.993), with a 95% confidence interval of −13.85 to 13.97. These results indicate that the diagnostic performance was comparable between the two methods.

This study compared four VSRAD scores and three volumetric measurements derived from the standard 3D T1-weighted MRI (T1w_3D) and the proposed SynthSR-based method (T1w_2DSR) using a multipronged statistical approach. The Wilcoxon signed-rank test revealed statistically significant differences across all four VSRAD scores and the three segmented brain volumes. The proposed method generally yielded larger volumetric estimates, even after Bonferroni correction. These differences indicate systematic biases, likely attributable to the SynthSR algorithm. Robust Bland–Altman analysis provided further insight by showing small median differences in most scores and volumes (30–32). These discrepancies may reflect the generative behavior of SynthSR. SynthSR likely compensated for thinned gray matter in cases of severe parenchymal atrophy, thereby increasing the apparent volume in the T1w_2DSR images. This effect was particularly evident in cases with ischemic lesions or cortical atrophy, where segmentation of GM, WM, and CSF failed. These cases were characterized by severe brain atrophy due to aging and enlarged ventricles (Figure 3) (42, 43). Furthermore, GM and WM volumes were larger in T1w_2DSR than in T1w_3D. The overall stronger degree of atrophy may be attributed to the advanced age of our cohort (mean age 83.5 years), compared to the VSRAD reference data cohort, which consisted of healthy subjects aged 54–86 years (44, 45).

Because Spearman’s ρ captures monotonic associations, several metrics such as Score 3 “Ratio” and the total brain volume demonstrated strong correlations, underscoring the reliability of SynthSR. Other indices showed moderate correlations. Our findings are generally consistent with the previous validation studies of SynthSR, particularly regarding correlation strength (46–48). Iglesias et al. demonstrated compatibility with morphometric properties derived from FreeSurfer and reported minimal bias across various brain structures (14). From a technical perspective, the use of Spearman’s rank correlation and robust agreement analysis follows recent recommendations for comprehensive evaluation of segmentation pipelines (49, 50).

The ROC curve analysis demonstrated high diagnostic accuracy for both T1w_3D (AUC = 0.90, Youden Index = 0.72) and T1w_2DSR (AUC = 0.96, Youden Index = 0.86). These findings suggest that both methods effectively differentiate between AD and normal cases. Although the proposed method (T1w_2DSR) showed numerically higher diagnostic metrics, this difference was not statistically significant (DeLong test: Z = 0.009, p = 0.99). Importantly, in clinical practice, Score 1 “Severity”—which reflects the degree of gray matter atrophy in the medial temporal lobe—is a critical index for early AD detection and monitoring. The consistently high AUC and favorable diagnostic characteristics of the proposed method support its use as a reliable alternative when standard 3D imaging is not feasible. Notably, SynthSR has been employed in low-field MRI settings (15, 17) and has also been used for segmentation correction in lesioned or incomplete scans (50, 51). These prior studies, along with applications in predicting thrombectomy outcomes (19) and tracking cortical lesions in traumatic brain injury (52), support the utility of SynthSR-enhanced reconstructions not only for advanced neuroimaging workflows but also for routine clinical scenarios, especially when standard 3D acquisitions are unavailable or degraded.

In DARTEL-based analyses such as VSRAD, the MRI data quality is critical, as factors such as head motion can significantly influence the results (53–56). In the present study, the quality of image data acquisition likely affected the results. Various strategies have been proposed to overcome challenges in acquiring 3D T1-weighted images. For example, Katayama et al. suggested using scout images for positioning to shorten scan time, although they reported significant differences in gray matter volume (55, 57, 58).

We suggest that SynthSR-enhanced volumetry is a viable alternative to standard 3D acquisitions in both clinical and research settings, particularly when scan time constraints or motion artifacts are of concern. In our cohort, T1w_2DSR reduced scan time to less than half (105 vs. 232 s). This reduction likely helped mitigate motion artifacts. Nevertheless, five T1w_3D cases had to be excluded due to motion artifacts. It is worth noting that our cohort was older than the typical target population for VSRAD, and the absence of parietal regions in the 2D images precluded intracranial volume comparisons. While this limitation restricts direct volume ratio analyses, Z-score-based assessments, such as those used in VSRAD, may still benefit from SynthSR-derived inputs. Moreover, this approach may also be applicable for longitudinal monitoring of individual patients. Previous studies have reported that SynthSR can improve the quality of low-field or heterogeneous MRI datasets, and it has been applied for segmentation correction in lesioned brains, harmonization across scanner types, and enhancement of legacy datasets (15, 17, 19). However, to our knowledge, its direct application to generate 3D-equivalent volumes from 2D T1-weighted inputs for VSRAD analysis in Alzheimer’s disease has not been systematically evaluated. The innovative aspect of this study lies in leveraging paired 2D and 3D acquisitions obtained on the same day to validate the feasibility of SynthSR-based reconstructions specifically for VSRAD indices, which are widely used in clinical practice in Japan. This approach demonstrates that retrospective 2D scans can be repurposed for quantitative dementia assessment, potentially expanding access to VSRAD analysis in settings where 3D scans are unavailable or degraded.

The greatest strength of this study lies in its comprehensive and robust statistical evaluation framework. However, several limitations must also be acknowledged. These include the lack of manual segmentation as ground truth, the generally older age of subjects, and the use of single-center data. Another important limitation of this study is that our cohort consisted predominantly of elderly individuals. This reflects the real-world demographics of patients undergoing dementia screening in Japan, where most individuals referred for MRI are already in advanced age. While this population is clinically relevant, the advanced age makes it challenging to fully disentangle age-related brain atrophy from Alzheimer’s disease–related neurodegeneration. Validation in younger subjects within the typical age range of AD onset would therefore be desirable. However, such a dataset was not available for the present retrospective single-center study. Future multicenter investigations that include younger cohorts and a broader age distribution will be essential to confirm the generalizability of our findings. Detailed clinical or biomarker characterization was not systematically available and was not the primary aim of this methodological validation study. Our primary aim was to assess comparability between standard and SynthSR-derived 3D volumes, independent of clinical diagnosis. Future studies should include multicenter validation across diverse populations, along with manual labeling to establish a reliable reference standard. Furthermore, although SynthSR revealed significant differences in many indices, the correlation coefficients remained relatively strong. This suggests that further optimization is warranted for specific brain structures and pathological conditions. Notably, the retrospective use of archival 2D images enabled by SynthSR may facilitate large-scale longitudinal studies. This approach holds promise for uncovering novel insights into neurodegeneration and its potentially modifiable risk factors.

This study focused on methodological validation rather than on diagnostic accuracy. Specifically, we evaluated whether 3D volumes generated using SynthSR from conventional 2D inputs yielded results comparable to standard 3D acquisition in a clinically relevant cohort of elderly individuals undergoing MRI for suspected cognitive decline. Considering that conventional 3D images are often difficult to obtain in this population, our findings suggest that SynthSR-based reconstruction may represent a practical alternative for VSRAD analysis in daily clinical practice.