Deep learning-based amyloid PET positivity classification model in the Alzheimer’s disease continuum by using 2-[¹⁸F]FDG PET

Alzheimer’s disease (AD) is characterized by the accumulation of β-amyloid (Aβ) and tau proteins. Aβ can be measured in humans with specific position emission tomography (PET) tracers or an examination of the cerebrospinal fluid (CSF) [1‐3]. Patients clinically diagnosed with mild cognitive impairment (MCI) or AD have been found Aβ-negative [4, 5], thus leading to a pathophysiology-based unbiased and descriptive amyloid, tau, and neurodegeneration classification [6, 7]. The aforementioned criterion considers the Alzheimer’s continuum only when the Aβ marker is positive. However, high cost and low availability make it difficult to conduct amyloid PET. Furthermore, the limitations of CSF examination can be attributed to its invasiveness, thus necessitating alternative ways to classification amyloid status.

2-deoxy-2-[fluorine-18]fluoro-D-glucose (2-[¹⁸F]FDG) is the most widely used PET tracer for measuring brain metabolism, which is related to neuronal activity [8]. Studies using 2-[¹⁸F]FDG PET images reported on lesser hypometabolism in the bilateral temporoparietal regions and hippocampus in Aβ-negative participants with MCI and AD, compared to their positive counterparts [4]. There had been several attempts to develop an amyloid status classification model using magnetic resonance imaging (MRI), neuropsychological, or laboratory tests [9‐11]. However, the performance was unsatisfactory. No attempts have been made using 2-[¹⁸F]FDG PET images.

Deep learning (DL) is the state-of-the-art mathematical algorithms that enable computers to automatically find patterns in large datasets. DL is studied in the medical imaging field for classification (diagnosis) [12, 13], predict prognosis [14, 15], detection [16, 17], and segmentation [18, 19]. A convolutional neural network (CNN), a subset of DL [20], is a DL model mimicking visual recognition concept which can extract features that reflect spatial relationships by applying non-linear convolutional filters. These spatial features pass through artificial neurons whose weights in the layers are properly set during training. Through this process, the model automatically learns hidden representative features from images and labels. 2-D-based deep learning architectures for 3-D medical images had been proposed: grid method (single montage image made by 16 images [21]), surface projection method (volumetric information projected onto a surface [22, 23]), and 2.5-D model (three axes (axial, coronal, and sagittal) images of brain volume [24]).

We aimed to construct a model that classifies amyloid PET positivity using 2-[¹⁸F]FDG PET that reflects brain metabolism and CNN architecture. In addition, we intended to validate the model using an independent external dataset with various diseases and classification probability analysis of submodules.

A sound understanding of amyloid pathology is critical for the accurate diagnosis and prognosis prediction in the Alzheimer’s disease continuum or in patients with dementia. The following two methods are currently available to examine amyloid pathology: amyloid PET and CSF analysis through lumbar puncture. However, amyloid PET is expensive and not universally available. Moreover, lumbar puncture is relatively invasive and poses potential side effects. Thus, the classification of amyloid pathology using long-established and widely distributed imaging techniques, such as MRI or 2-[¹⁸F]FDG PET, appears promising. Several studies have attempted to classify amyloid pathology by analyzing the cortical thickness from structural MRI (sMRI) [9, 10]. Nonetheless, sMRI has its limitations in early classification. This can be attributed to the occurrence of cortical thinning primarily at the late stage of the disease. Moreover, it does not provide any functional information. Glucose metabolism changes before a change in the cortical thickness in AD [37]. Hence, we assumed that 2-[¹⁸F]FDG PET would better classify amyloid pathology. However, the use of 2-[¹⁸F]FDG PET images for amyloid classification is rare. Thus, we developed a DL model that determines amyloid pathology from 2-[¹⁸F]FDG PET imaging data.

We found a lower glucose metabolism in the Aβ-PET-positive participants in the bilateral angular, supramarginal, precuneus, middle and posterior cingulate, inferior and middle temporal, parahippocampal, and fusiform gyri, consistent with previous reports (Fig. 4a) [4]. However, we recorded a higher glucose metabolism in the bilateral paracentral lobules and pre- and postcentral gyri in the Aβ-positive participants. According to Braak staging, the aforementioned regions are reportedly preserved during the progress of Aβ pathology [38]. These findings support the possibility of its classification power and form the basis of model interpretation.

Despite these possibilities, the datasets were extremely small for building a 3-D DL model. However, the model could extract more spatial relationships while considering the coronal and sagittal images, in addition to the axial images. Therefore, we designed a 2.5-D DL model for completely considering the 3-D spatial relationships. The CNN DL model can be usually explained using CAM for visualization. However, the last layer of each submodule gets replaced with a global average pooling layer during CAM. Moreover, the values from the submodules passed the fully connected layer for the final classification. This in turn leads to the disappearance of the location. Thus, the analysis would provide limited spatial information in 2.5-D. Therefore, we developed a 2.5-D method that analyzed the prediction values from the submodules to understand the basis of this decision.

We trained and internally validated the combined internal datasets. The internal validation results were impressive, with AUC values, accuracy, sensitivity, specificity, and F1 score of 0.811, 0.733, 0.678, 0.785, and 0.709, respectively. We encounter several diseases characterized by memory impairment in clinical practice. In contrast, the ANDI and KBASE datasets only contained the cognitive unimpaired and Alzheimer’s disease continuum. Thus, we externally validated our proposed model using a dataset, including various neurological diseases that cause memory impairment. The external validation dataset included clinically diagnosed AD, Lewy body dementia, epileptic cognitive disorder, normal pressure hydrocephalus, frontotemporal dementia, and vascular dementia. In addition, it included various clinical stages, such as one, 75, and 24 patients with cognitive unimpaired, MCI, and dementia, similar to the clinical situation. Moreover, the application of the proposed model to the external validation dataset also resulted in a high AUC value, accuracy, sensitivity, specificity, and F1 score of 0.798, 0.690, 0.768, 0.612, and 0.712, respectively. Furthermore, when we performed external validation with the model shown the best AUC value (0.864) on the internal validation dataset, the AUC value, accuracy, sensitivity, specificity, and F1 score was 0.860, 0.770, 0.800, 0.740, and 0.780, respectively. The performance of our model is comparable to that of the model based on longitudinal sMRI (AUC value was 0.86) [9]. However, it is better than that of the model which used patient demographics and sMRI (AUC value was 0.79) [10]. This enhanced performance could be attributed to the changes in glucose metabolism before cortical atrophy. Thus, our model that only used one-time 2-[¹⁸F]FDG PET images could be effectively used for classifying amyloid PET positivity in memory disorder clinics.

The performances by subgroups in the internal validation resulted in a high AUC of 0.717, 0.757, and 0.816 in the CU, MCI, and AD participants, respectively. Our model outperformed conventional models by producing comparatively better AUC in AD subgroup that indicated many significant differences within the group on voxel-wise analysis. Additionally, it resulted in a good AUC in CU subgroup that indicated a few significant differences within the group on voxel-wise analysis. Moreover, the application of the proposed model to the external validation dataset also resulted in a high AUC of 0.798 and 0.787 in the MCI and demented participants, respectively. These results were achieved because the external validation dataset contained many MCI and demented participants data. Additionally, the result of the internal validation in the Alzheimer’s Continuum was impressive, but the results for when this model was applied on the external validation dataset with various clinical diagnosis were impressive as well. This reflects the applicability in clinical practice.

When our model was trained and internally validated in the ADNI dataset and externally validated in the KBASE dataset, the model achieved the AUC values, accuracy, sensitivity, specificity, and F1 score were 0.769, 0.698, 0.702, 0.694, and 0.719, respectively. These results suggested that the performance was comparatively decent than when we used the ADNI and KBASE dataset for training and internal validation (Additional file 1: Fig. 4 and Additional file 1: Table 1). The reasons of the differences were suggested as follows: First, the ADNI and KBASE had different patients’ composition. Within the ADNI dataset, MCI participants were the most common, and 55% of the participants had amyloid PET-positive scan, while, within the KBASE, CU participants were the most common, and 34% of the participants had amyloid PET-positive scan. Second, the two datasets were obtained under different ethnicities, different amyloid tracers, and various PET manufacturers. Although we did spatial normalization with ethnicity specific template in the preprocessing pipeline, there may be remained differences. However, from the other perspective, the above results were also affirmative. The model with the best AUC value (0.800) in the internal validation showed good AUC values of 0.745 and 0.984 in MCI and AD, respectively, which also showed good AUC values of 0.837 and 0.721 in MCI and AD in the external validation, respectively. These results were encouraging since one could classify amyloid PET positivity status from the dataset obtained regardless of the ethnicities, tracers, and MRI manufacturers.

On internal and external validation, 10–15% of false negative and false positive cases were found. According to Alzheimer’s disease biomarker curve [39], amyloid deposition preceded the hypometabolism, which could explain the false negative cases. There are two speculation for comprehend false positive cases. First, approximately 10% of false negative cases are due to sensitivity of amyloid PET [40, 41]. Second, other cause of dementia, including Lewy body dementia, could be shown similar hypometabolism pattern [42].

The common region between the analysis of prediction value from the submodules and voxel-wise analysis included the right superior parietal gyrus in the external validation dataset (Fig. 6). However, we observed additional regions with different metabolism in the training and internal validation datasets that only included AD (Fig. 5). The above-mentioned difference could be attributed to the characteristics of the patient population, including only AD or other diseases. Therefore, our model might have extracted these hidden features and used them for the classification.

Our study had some limitations. First, we used several datasets that used different protocols and amyloid tracers for data collection. Moreover, these datasets included patients of various ethnicities. However, their characteristics suggest the general applicability of our model regardless of the aforementioned variabilities. Second, amyloid pathology could be detected by CSF Aβ₄₂ or amyloid PET studies; most cases show concordant results in the Alzheimer’s disease continuum, whereas 6–21% show discordant results [43]. Both results should be considered for more accurate amyloid pathology; however, CSF Aβ₄₂ was not available in all datasets. Especially in the ADNI dataset, only 66 participants underwent sMRI, 2-[¹⁸F]FDG PET, amyloid PET, and CSF Aβ₄₂. We chose amyloid PET only because deep learning requires a large amount of data. Third, the DL framework had a black box limitation. Thus, the model visualization method, such as CAM, can further explain our classification model. We could not apply CAM because of the limitation of available datasets. In contrast, we used three axes images (2.5-D) as the inputs to improve the classification accuracy. The development of a 3-D DL model from larger datasets would facilitate the interpretation through CAM. Fourth, we suggested a novel method for the brief understanding of our model’s decision. After the t test between the amyloid PET-positive and PET-negative participants on SPSS, we obtained 20 slices in the sagittal plane, 27 slices in the coronal plane, and 24 slices in the coronal plane. We considered the (x, y, z) point as statistically significant when the ‘x’th, ‘y’th, and ‘z’th planes of the sagittal, coronal, and axial axes, respectively, showed substantial differences; a total of 12,960 (20×27×24) points were plotted in the MNI space. However, our method suffers from a limitation due to the following assumption. If the (x,y,z) point is significant, the ‘x’ th plane in the sagittal, ‘y’th plane in the coronal, and ‘z’th plane in the axial could be significant; however, the opposite was not assumed. Therefore, many unrelated points were plotted in the MNI space; ROIs clustered with more than 50 points were extracted to remove these unrelated points. Further investigation is required to accurately understand the model’s decision. Fifth, we did not include other biomarkers or neuropsychological test results. An incorporation of these features might have improved the performance. However, we tried to build a classification model by only using images, which can be applied more generally.

In conclusion, we proposed a DL model that can classify the amyloid PET positivity from 2-[¹⁸F]FDG PET imaging and demonstrated its high performance across an external test dataset. A large-scale external validation of multi-institutional data, model calibration, and optimization of sensitivity needs to be incorporated into the clinical workflow. The aforementioned model can serve as an important decision supporting tool to aid clinicians while classifying amyloid PET positivity.

The proposed model based on the 2-[¹⁸F]FDG PET imaging data and a DL framework might successfully classify amyloid PET positivity in clinical practice, without performing amyloid PET, which have limited accessibility.