Alzheimer’s disease (AD) is a neurodegenerative disorder characterized by progressive cognitive decline, with pathological hallmarks including amyloid-beta (Aβ) plaque deposition and neurofibrillary tangles formation. The prevalence of AD in individuals over 60 years old is estimated to be 3.7% (1), making it one of the leading causes of cognitive impairment among the elderly, posing a significant burden on individuals, families, and society (2). In the absence of effective prevention and treatment strategies, the number of AD patients is projected to rise dramatically to 13.8 million by 2060 (3). This underscores the increasing need for early identification of high-risk individuals and early diagnosis of AD (4).

Aβ plaque deposition is a hallmark pathological feature of AD and a necessary condition in the AT(N) diagnostic framework. As a non-invasive molecular imaging technique, PET/CT imaging is capable of localizing and quantifying specific biomarkers, thereby providing significant evaluative value for the diagnosis of AD, as well as for assessing disease progression and prognosis (5, 6). However, Aβ PET/CT positivity is not exclusive to AD, as some NAD patients also exhibit similar imaging features, complicating diagnosis and treatment strategies.

Radiomics technology enables the rapid generation of high-throughput imaging features, capturing pathological changes that are imperceptible to the naked eye and may be closely associated with cellular or molecular alterations (5, 7). Currently, radiomics has been applied to AD PET imaging to explore early diagnosis and differential diagnosis through various molecular imaging approaches. Previous studies have demonstrated that Aβ PET radiomic features can serve as novel biomarkers for the clinical application of AD and mild cognitive impairment (MCI) (8). In contrast, the widely used conventional SUVr method is susceptible to multiple factors, including tracer type and analytical processes. Numerous studies have reported that PET radiomics significantly outperforms the SUVr method in diagnostic performance (9–12), suggesting that radiomics may provide a more precise and efficient approach for Aβ PET image analysis, reducing the ambiguities associated with visual interpretation. This study aims to further investigate the potential value of radiomics in the diagnosis and differentiation of AD.

Alzheimer’s disease (AD) is a severe neurodegenerative disorder, and its diagnosis remains a major challenge in the medical field. The application of Aβ (amyloid-beta) PET/CT imaging tracers has brought significant promise for early disease detection. These tracers enable both qualitative and quantitative assessments of AD by capturing the distribution and concentration of Aβ deposition in the brain. However, the clinical heterogeneity and complexity of AD result in overlapping clinical symptoms and imaging features with non-AD patients, which complicates accurate diagnosis, prognosis, and targeted treatment (20, 21). Therefore, developing more advanced biomarkers and models to improve metabolic information and predictive ability are still crucial for distinguishing AD from non-AD (9).

In recent years, the emergence of radiomics has provided an advanced approach to disease diagnosis. By extracting high-throughput imaging features, such as texture, morphology, and grayscale distribution, radiomics captures both low-order and high-order statistical attributes of imaging data, making it particularly suitable for complex, heterogeneous, and multifaceted diseases. This has improved the accuracy of clinical diagnosis, prognosis, and prediction, and has increasingly attracted attention in the study of brain diseases (22, 23). The increasing application of machine learning has further strengthened radiomics analysis, allowing for more comprehensive models to handle high-dimensional and multi-variable data compared to traditional methods. These models facilitate disease prediction and accurate classification of overlapping symptoms, thereby aiding clinical decision-making (24). More recently, deep learning has been increasingly used for neuroimaging tasks involving classification and prediction. Compared to traditional approaches, deep learning models have the advantage of minimal inference time while eliminating the need for complex image preprocessing steps (25). These advancements provide an optimal solution for AD diagnosis and differentiation and have demonstrated promising classification performance (26). The integration of radiomics and machine learning offers a vast research potential for novel imaging data analysis methods.

This study utilized Aβ PET/CT imaging data to extract 15 radiomic features, identifying four features with strong correlations to AD diagnosis, primarily located in the frontal, temporal, and parietal lobes. The most significant features included original_firstorder_Skewness and original_glcm_ClusterShade. original_firstorder_Skewness reflects the skewness of pixel intensity distribution, indicating asymmetrical metabolic or pathological deposition patterns in specific brain regions. Meanwhile, original_glcm_ClusterShade, derived from the gray-level co-occurrence matrix, quantifies texture complexity and captures microstructural variations in brain gray matter. This complements the limitations of the SUVR single quantitative value (10). And has a significant association with the clinical diagnosis than SUVR (27). The distribution of these radiomic features closely aligned with findings from the SUVr method (involving the frontal, temporal, and parietal lobes), which is also consistent with previous studies (8, 28, 29), highlighting the critical role of these brain regions in AD diagnosis. Additionally, in the SUVr method, moder emerged as a particularly strong diagnostic feature. As a quantitative metric, moder reflects the mode signal intensity within a specific brain region, providing a detailed characterization of regional Aβ pathology distribution.

Comparing radiomics and SUVr methods, the Radiomics_r model demonstrated significantly superior diagnostic performance based on ROC curves and DCA decision curve analysis, achieving an AUC of 0.89 and an accuracy of 0.88. These findings align well with literature reports (10, 11, 30, 31), where AUC values for AD differentiation typically approach 0.9, and PET-based classification of AD versus normal controls achieves accuracy rates of 80–90% (32, 33). In contrast, the SUVr model exhibited lower AUC, sensitivity, specificity, and accuracy, with specificity being particularly low (0.45). This is consistent with prior findings, as Aβ PET has been reported to have relatively low specificity (below 60%) (34). Similarly, a meta-analysis by Morris et al. (35) found that while Aβ PET is highly sensitive for detecting AD and demonstrates good overall diagnostic efficacy, its specificity remains moderate. This is largely due to a high false-positive rate, as other types of dementia, such as dementia with Lewy bodies (DLB) and frontotemporal dementia (FTD), may also present AD-like pathological changes. Additionally, SUVr values often fall within a research-defined range, with whole-brain cortical SUVr thresholds typically between 1.2 and 1.5 (36). Chanisa et al. (37) observed that β-amyloid plaques tend to spread from the cerebral cortex to the cingulate and precuneus regions, with the highest 11C-PIB deposition in the anterior and posterior cingulate gyri. Their study also suggested that regional SUVr-based diagnosis provides greater sensitivity and specificity than whole-brain SUVr, with regional SUVr cut-off values between 1.46 and 1.81. This highlights the substantial overlap between different pathological conditions, which our study aims to address by distinguishing Aβ PET-positive lesions in AD and NAD patients. Our results confirm the irreplaceable advantage of radiomics in overcoming this challenge.

Interestingly, the combined SUVr + Radiomics_r model exhibited lower diagnostic performance than the Radiomics_r model alone. This may be attributed to fundamental differences in how SUVr and Radiomics_r features characterize imaging data, suggesting that radiomics not only enhances the differentiation between AD and NAD patients but also provides a more precise capture of key disease characteristics.

It is worth noting that despite the inferior overall performance of the SUVr model compared to radiomics, there remains potential for improving its diagnostic capability. In cases where full radiomics analysis is not feasible, selecting moder values from the frontal, temporal, and parietal lobes as supplementary indicators could enhance the reliability of SUVr-based diagnosis. This finding provides new insights for optimizing traditional methods.

This study has several limitations. First, radiomics analysis relies on large, high-quality datasets, and our study was limited by a relatively small sample size and single model fitting. Larger datasets are needed to validate the generalizability of our findings and assess results stability through cross-validation. Second, we only extracted PET-based features, as CT and PET/CT standardized brain templates were unavailable, preventing the extraction of multimodal features. Furthermore, MRI imaging data were not incorporated, despite its importance in structural brain analysis. The inclusion of MRI data in future studies could provide a more comprehensive assessment of radiomic features. Additionally, variations in tracer affinity, pharmacokinetics, region of interest selection, reference regions, and imaging acquisition parameters (e.g., tracer dosage, scan timing, and image reconstruction techniques) may lead to inconsistencies across studies, emphasizing the need for standardized quantitative methodologies to ensure reproducibility across different research centers.

This study demonstrates the significant potential of radiomics and machine learning in Aβ PET/CT-based AD diagnosis and differentiation. Radiomic features derived from Aβ PET imaging could serve as novel neuroimaging biomarkers for clinical applications in AD. Additionally, Aβ PET imaging enables continuous monitoring of brain Aβ burden dynamics, establishing correlations with disease progression and facilitating treatment guidance and therapeutic efficacy evaluation.