Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD.

In our study, we analyzed a total of 2,298 FDG-PET imaging studies of 1,045 patients obtained from ADNI. The datasets contained images of subjects of different ages. In ADNI 1, the subjects were grouped into three classes: CN, MCI, and AD. However, in ADNI 2/GO, the MCI stage was subdivided into EMCI and LMCI. To be classified as CN, subjects must have no memory complaints and be non-demented. To be classified as having MCI, subjects must have a Mini-Mental State Examination (MMSE) score between 24 and 30; the activities of daily living must be preserved, and dementia must be absent. Finally, to be classified as having AD, subjects must be clinically diagnosed as such, with an MMSE score between 20 and 26 (Jack Jr et al., 2008). Demographic information on our dataset is presented in Table 1. A total of 80% of the data (1,851 imaging studies, 598 patients) were used for model training. The remaining 20% (447 imaging studies; no repeat studies of the same subjects in the test set) were used for model testing, from which an additional test set (74 imaging studies for AD vs. MCI vs. CN classification and 45 imaging studies for EMCI vs. LMCI classification) was selected for validation by professional radiologists.

For the purpose of eliminating differences between images acquired from various systems, FDG-PET images in ADNI have undergone a series of preprocessing steps, intensity normalization, and conversion to a uniform isotropic resolution of 8 mm full width at half maximum. We selected the processed images from ADNI; our method does not require any specific pre-defined ROI or VOI as traditional machine learning methods do. All 3D images were resampled to a size of 160 × 160 × 96; we treated the images as a series of 2D slices and removed slices with all-zero intensity on both sides, then divided the image into four groups of slices at equal intervals, with each group containing 23 slices. All processing steps were conducted in Python (version 3.8) using the packages scipy (http://www.scipy.org) and numpy (https://numpy.org/). Figure 1 shows a single slice, viewed on three planes.

Despite the good learning ability of deep neural networks, they are easy to overfit on small datasets and their training is also time-consuming. The BLS is a lightweight network with a broad structure proposed by Chen and Liu (2017). The inspiration for its design comes from a random vector functional link neural network (RVFLLNN) (Pao and Takefuji, 1992; Chu et al., 2019; Gong et al., 2021). It can obtain a globally optimal solution using a ridge regression algorithm during training. without an iterative backpropagation process, meaning that its training is fast and efficient. The detailed description of the BLS is illustrated in the Supplementary material. Based on the BLS, we propose a broad network-based model for early diagnosis of AD (BLADNet) through PET imaging of the brain.

Figure 2 illustrates the overall architecture of BLADNet. The framework consists of two stages. In the first stage, we use a 2D CNN for automated feature learning from each group of slices rather than directly using a 3D CNN, which reduces the number of parameters to be learned. In the second stage, the features extracted from each group in the previous step are concatenated to form a compact sequence feature; then, the Extreme Broad Learning System (EBLS), based on a broad neural network, is used to enhance the features from 2D CNN and carry out the final classification. A detailed description of the EBLS is provided in the Supplementary material¹. Our model was developed in Python using the packages numpy and pytorch (https://pytorch.org/, version 1.7.1). All experiments were conducted using a computer with a Linux operating system (Ubuntu 18.04). The computer was equipped with a CPU (Intel(R) Core (TM) i9-9980XE, 3.00 GHz), 64 GB of DDR4 SDRAM, and GPU (GeForce RTX 3080) with CUDA Version 11.2 and cuDNN Version 9.1.85.

We performed the experiments of AD vs. MCI vs. CN classification as in previous studies on data from ADNI 1, and also performed additional experiments of refined classification between EMCI and LMCI on data from ADNI 2/GO. All data were randomly shuffled before being spli into the training set and test set. In all experiments, we used 80% of the samples for training and 20% of the samples for testing. In the experiment, we regard each category as positive samples respectively, the rest as negative samples, and then calculate metrics. We used accuracy, sensitivity, and specificity as metrics to evaluate classification performance. All metrics were calculated under a default threshold value of 0.5. We also plotted the ROC curve of all experiments and calculated the corresponding AUC.

Two board-certified professional radiologists working in a department of brain imaging and nuclear medicine (radiologist 1: HLZ, with 8 years of experience in brain imaging reading for AD diagnosis; radiologist 2: HHW, with 6 years of experience in brain imaging reading for AD diagnosis) were asked to give their diagnostic impressions of a dataset that was not used for model training. For each case, the radiologists were provided with the patient's age, gender, and MMSE score as additional information for validation. To validate the performance of the proposed model and the professional readings of radiologists, we compared the performance of our proposed model with that of the radiologists' interpretations. The main steps of the experiment are shown in Figure 3.

The prediction results of the broad network-based model are shown in Table 2. For classification of AD, MCI, and CN samples, sensitivity was 92.16 (94 of 102), 89.34 (109 of 122), and 95.16% (118 of 124), respectively; specificity was 97.56 (240 of 246), 95.58 (216 of 226), and 95.09% (213 of 224), respectively; and precision was 94.00 (94 of 100), 91.60 (109 of 119), and 91.47% (118 of 129), respectively. The ROC curves of our model, trained on 80% of the ADNI data and tested on the remaining 20%, are shown in Figure 4. The AUC in prediction of AD, MCI, and CN was 0.97, 0.98, and 0.99, respectively. The AUC for CN was the highest, indicating that our model can distinguish healthy subjects from patients with AD/MCI.

The results for EMCI and LMCI prediction are shown in Table 3. In this experiment, we treated LMCI as the positive class and EMCI as the negative class. Sensitivity was 81.63% (40 of 49) and specificity was 85.19% (46 of 54). Similar to the AD vs. MCI vs. CN experiment, the specificity of the model was much higher than the sensitivity, indicating that our model was better than radiologists at identifying healthy subjects.

We used the t-SNE algorithm to reduce the dimensionality of the features extracted from the convolutional network and projected them into a two-dimensional space for visualization. As shown in Figure 5A, for the AD vs. MCI vs. CN experiment, there were obvious boundaries between the three categories. Moreover, only a few samples from other categories were scattered within the CN category, indicating that the model has a better screening ability for healthy cases than for patients. Similarly, as shown in Figure 5B, for EMCI and LMCI classification, the model divided the samples very successfully into two clusters. Although a few cases were mixed in the junction of the two clusters, which indicates that there is a transition stage from EMCI to LMCI, our model could distinguish the two stages well.

Recently, a substantial amount of work has been carried out exploring the application of machine learning approaches to AD prediction using brain imaging. Most of these studies have used structural imaging of the brain, with few studies using functional imaging, specifically 18F-FDG-PET. Some researchers have attempted to analyze 18F-FDG-PET for AD predictions, but these studies have yielded limited success (Liu et al., 2018; Lu et al., 2018; Pan et al., 2018; Ding et al., 2019; Huang et al., 2019; Hamdi et al., 2022). Tables 4, 5 summarize state-of-the-art deep learning methods for prediction of AD using 18F-FDG-PET imaging. Most of the methods investigated can only discriminate AD from CN or MCI from CN, while our method can classify patients at different stages of AD with higher sensitivity and specificity. In addition, compared with these methods, we used a larger test set in our experiments, which demonstrates the superior generalization ability of our method.

As shown in Table 2, two radiologists gave their interpretations of a test set. For radiologist 1, the sensitivity results for MCI, AD, and CN were 51.85 (14 of 27), 29.41 (5 of 17), and 46.67% (14 of 30), respectively; the specificity results were 57.45 (27 of 47), 80.70 (46 of 57), and 77.27% (34 of 44), respectively; and the precision results were 41.76 (14 of 34), 34.25 (5 of 16), and 58.33% (14 of 24), respectively. For radiologist 2, sensitivity for MCI, AD, and CN was 37.04 (10 of 27), 35.29 (6 of 17), and 46.67% (14 of 30), respectively; specificity was 72.34 (34 of 47), 63.16 (36 of 57), and 77.27% (34 of 44), respectively; and precision was 43.48 (10 of 23), 22.22 (6 of 27), and 58.33% (14 of 24), respectively. It can be observed that the prediction results of our proposed model were better than those of the radiologist, which indicates that the model was able to find lesions that were difficult to observe with the naked eye. It is also worth noting that although the two radiologists obtained the same results in their evaluations of healthy cases, patients with MCI and AD were difficult to evaluate.

Table 3 reports reader performance on prediction of EMCI vs. LMCI. For radiologist 1, the results in terms of sensitivity, specificity, and precision were 84.00 (21 of 25), 25.00 (5 of 20), and 58.33% (21 of 36), respectively. For radiologist 2, the results were 76.00 (19 of 25), 30.00 (6 of 20), and 58.33% (19 of 33), respectively. Although radiologists had higher sensitivity in this scenario, their specificity was very low; this is because radiologists tend to predict cases as LMCI. In contrast, our model was able to achieve high specificity under high sensitivity.