Automated Medical Diagnosis of Alzheimer´s Disease Using an Efficient Net Convolutional Neural Network

Alzheimer’s disease (AD), the most prevalent kind of dementia, has no recognized disease-modifying therapy to date. It is characterized by a silent onset, in which AD gradually advances over a number of years before any clinical signs appear [1]. At least fifty million people worldwide are thought to be affected by AD and other forms of dementia [2]. As of 2022, there are 7 million AD patients in the USA, and 14 million individuals are predicted to be affected by the disease by 2050 [3]. For various forms of dementia, such as AD or mild cognitive impairment (MCI), it is essential to investigate novel early diagnostic techniques to ensure proper treatment and halt the progression of the illness. Between healthy cognitive function and AD lies a condition known as MCI. A person with MCI presents cognitive impairment but is nevertheless capable of conducting daily activities. MCI affects around one-fifth of the population over 65 years old, and about one-third of them will develop AD within three to five years [3]. Subjects with MCI will either develop AD or remain stable. Only autopsy can certify an AD diagnosis [4]. Nonetheless, biological and functional brain problems linked to AD may be examined and assessed using magnetic resonance imaging (MRI), often utilized in clinical practice and recognized as a helpful technique to identify the course of AD [4, 5].

Computer-aided machine learning (ML) techniques offer a systematic means to create complex, automatic classification models to manage large amounts of data and can find intricate and subtle patterns. Support vector machines (SVMs), logistic regression (LR), and SVM-recursive feature elimination are ML pattern analysis techniques that have been proved successful in AD detection [6]. However, automated diagnostic ML models for neuropsychiatric disorders based on SVMs require hand-made features because they cannot pull out adaptive features [6]. When utilizing these ML approaches for classification, the architectural design must be established. In general, four phases are required: feature extraction, feature selection, dimension reduction, and implementation of a classification method. Additionally, this process must be optimized, and specialists need to be involved at every level [7]. The use of deep learning (DL) models to predict different AD stages was made possible by the growing GPU processing power. The field of ML known as DL simulates how the human brain finds complicated patterns. DL techniques like convolutional neural network (CNN) and sparse autoencoders [8‐15] have recently surpassed statistical ML techniques. The use of CNNs has rapidly spread into a variety of domains, beginning with the outstanding performance of AlexNet in ImageNet large-scale visual recognition challenge [16] and then expanded into medical image analysis, starting with 2D images, such as chest X-rays [17], and then progressing onto 3D images, including MRI. End-to-end learning (E2EL) is the central principle of DL. A key advantage of E2EL is that it simultaneously improves every stage of the processing pipeline, potentially resulting in an optimum performance [18].

For the analysis of MRI scans, Oh et al. [19] suggested an end-to-end hierarchy extending between 1 to 4 levels. Feature selection and extraction are carried out manually at Level 1 [20, 21]. Level 2 involves either the segmentation of 3D data into ROI or their conversion into 2D slices, following by their use as input to train the DL model, that can be either self-designed DL architecture or a CNN-based pertained transfer learning (TL) architecture like ResNet[12], deep ResNet [15], CaffeNet [22], DenseNet [23], etc.

According to recent research TL, which allows successful DL training even with limited data, is becoming a quite popular area of DL [24, 25]. TL is compared to human behavior as it may use the acquired information to solve new challenging situations.

When comparing stable MCI (sMCI) vs. progressive MCI (pMCI) or AD vs. sMCI classifications, it is simple enough to classify AD vs. CN, as there is a clear difference between the brain anatomy of AD and CN subjects and a sufficient availability of MRI scans [26]. Numerous studies [22, 27, 28] have also used local TL as a basis for this assumption. Local TL implies the classification of (sMCI, pMCI) or (sMCI, AD) subjects using the learning of the classifier that has been implemented to categorize (AD, CN) subjects. However, no studies using 3D pertained TL architecture have been found.

Preprocessed 3D MRI scans are fed into DL networks at Level 3. MRI scans must be preprocessed for any analytical analysis method to be effective [29]. During preprocessing, methods for noise removal, inhomogeneity correction, brain extraction, registration, leveling, and flattening are used to improve the quality of the image and make the architecture and brightness patterns consistent. A 3D MRI scan from the device is directly relayed as input into DL networks at Level 4; however, the authors are not aware of any studies utilizing this level.

Mehmood et al. extracted gray matter (GM) tissue from MRI scans and fed it into the VGG-19 architecture to classify different AD phases [30]. A multi-class categorization of AD and its related stages using rs-fMRI and ResNet18 was done by Ramzan et al. [31]. During preprocessing, several studies [15, 32] segment the GM area, which is there after used as input for CNNs. Through the use of 3D-stacked Convolutional Autoencoders and MRI data, Hosseini-Asl et al. [33], by utilizing a Level 3 approach, reported the first effective use of a volumetric CNN-based architecture to classify AD and CN subjects.

Multimodal DL techniques [34‐36] combine inputs from different data sources to better understand the structure and function of the brain by using several biological and physical properties to boost the classification accuracy of AD stages. Due to the limitations of the different resolutions, sheer number of dimensions, heterogeneous data, and limited sample sizes, multimodal DL techniques are particularly difficult to deploy at Level 3 learning [37]. In addition, we found that studies utilizing Level 2 learning mainly employed multimodal DL approaches. A unique deep neural network (DNN)-based approach was presented by Lu et al. [38] with multi modalities FDG-PET and T1-MRI to classify sMCI and pMCI subjects.Song et al. [39] created the "GM-PET" fused modality by combining the GM tissue region of MRI scans with FDG-PET images utilizing mask encoding and registration to aid in the diagnosis of AD. Only a small number of investigations utilized Level 3 architecture. The AD-NET was presented by Gao et al. [28], with the pre-training model performing the dual functions of extracting and transferring features as well as learning and transferring information. VoxCNN, based on deep 3D CNN, was proposed by Korolev et al. [40] for AD early detection. A multimodal DL framework based on multi-task CNN and 3D DenseNet was proposed by Liu et al. [41] for concurrent hippocampus segmentation and AD classification.

Literature demonstrates that early detection of AD is essential for the patient to obtain maximum benefit. As of right now, the highest accuracy level for this task utilizing either E2EL, local TL, or a CNN-based 2D transfer learning or ROI segmentation strategy is 86.30% [42]. Therefore, to enhance accuracy as well as the generalization capability of the model, we propose a fusion of E2EL and TL during the training phase of the model. We trained the EfficientNet-b0 CNN for the AD vs. sMCI binary classification task by transferring the learning of each fold of fivefold stratified cross-validation to its subsequent fold, and so on. In the first fold, the model was trained using E2EL. We also trained and evaluated an E2EL-based EfficientNet-b0 model for the multiclass AD vs. CN vs. sMCI classification task.

Preprocessed 3D T1W MRI scans were fed into the models. ANTsPyNet [43] was used to preprocess MRI scans, whereas the Medical Open Network for Artificial Intelligence (MONAI) was used to design and implement the models [44]. The entire implementation was done in PyTorch GPU utilizing Google CoLab Pro + . To conclude, the following are the main contributions made by this article:

The rest of this article is structured as follows: research methodology, datasets, input management, DL models, and experimental setup are described in “Materials & methods” section, followed by results and discussion in “Experimental setup” section, and conclusions in “Results and discussion” section.

We propose DL models for the classification of MRI scans of 1) AD and sMCI subjects, and 2) AD, sMCI, and CN subjects. For task one we used a novel E2EL and TL fusion approach, as shown in Fig. 1. During model training via fivefold stratified cross-validation, we trained the model from scratch in the first fold (E2EL), validated it, and used the final weights of the best epoch from fold 1 as the initial weights for fold 2 (TL). After training and validating the model in fold 2, we used the final weights of the best epoch from fold 2 as the initial weight for fold 3, and we repeated the same steps for the subsequent folds. For task 2, we used E2EL for all the folds. The model of the best epoch from each fold was used to assess external MRI scans to check for overfitting.

The whole implementation was completed using Google Colab Pro + [56], which was made available to the public in August 2021. Some of its most notable features include the ability to run in the background, early access to more powerful GPUs, and increased memory availability. Asynchronous data loading and multiprocessing are facilitated by GPUs. Despite not guaranteeing compatibility with a specific GPU, Colab Pro + does offer priority on the available options. Even with Pro + , GPU quality might decline after periods of intensive use. Pro + offers Tesla V100 or P100 NVIDIA Deep Learning GPU with CUDA support. The "High-RAM" option of Colab runtime met its objective by providing 52.8 GB RAM. Runtime support is supposed to be 24 h as stated in Colab's specs, yet we only got assistance for a maximum of 8 h. For this reason, we could not run all the folds at the same time in this setup, as finishing a fold with 50 epochs requires approximately 2 h. To implement a fivefold, stratified cross-validation, we first created five data sets (DATASETS 1–5) for training and validation, with the same class ratio as the original dataset across all folds. To train and validate Fold [n], we utilized Dataset [n]. The MRI scans were distributed as follows:

All scans were prepared according to the preprocessing workflow shown in Fig. 2. Following preprocessing, five datasets were constructed and used, respectively, in folds 1–5. All models were implemented using MONAI, an open-source, PyTorch-based DL healthcare imaging platform maintained by the community. MONAI provides built-in PyTorch support along with a set of core features suited for the medical imaging area. Two important hyperparameters, learning rate and number of epochs, were optimized using the random search technique. We trained and tested the model on three different combinations of learning rate and number of epochs and, based on the comparison of their performances as shown in Fig. 4, learning rate was adjusted to 0.0001 and number of epochs to 50 for the final fivefold stratified CV implementation.

Adam is the first widely used "adaptive optimizer." It was used with a learning rate of 0.0001 and a batch size of 2. In the experiment, most cases converged after approximately 50 epochs. In addition, we used the cross-entropy loss function and the area under the curve for receiver operating characteristics (ROCAUC) metric to optimize the model weights during training and to evaluate the discriminatory abilities of the model across classes. The implemented method for the classification of sMCI vs. AD is presented in Algorithm 2. As indicated in steps 2 and 3, we utilized MRI scans from DATASET [1] for training and validation. In steps 48 and 49, we initialized the training and validation loaders from DATSET [C] MRI scans to be analyzed in fold C.

In steps 43 to 46, we evaluated unseen MRI scans using the model with the best weights from each fold. We did not reset the model in step 50; conversely, we raised the counter to reflect the fold's rise since we intended to apply this fold's learning to the next fold. The same approach, with a few modifications, was applied to the AD, sMCI, and CN multiclassification task: i.e., there were three types of input MRI scans; the batch size and worker values were set to eight; the number of output classes was changed to three; and the number of epochs was increased to 100. The same process was used for testing, validation, and training inside each fold. After each fold, the model was reset to apply E2EL.

The research training outcomes were evaluated using the following five metrics, as they provide critical data for the thorough evaluation of models that have been put into use: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). Here, an AD patient is classified as TP or FN depending on whether they are placed in the AD group or not. Similarly, TN indicates the total number of individuals classified as presenting sMCI, and FP represents the total number of subjects not in the sMCI category. Accuracy (ACC) was defined as (TP + TN)/(TP + TN + FN + FP), while precision was defined as TP/(TP + FP). The positive sample prediction is more accurate when the precision is higher. Recall was defined as TP/(TP + FN), and the higher the recall rate, the more accurately the target sample may be predicted, and the less probable it is that a problematic sample will be missed. Precision and recall are often at odds with one another; hence, the F1-score is given as a composite statistic to balance their effects and evaluate classifiers more correctly. ROCAUC serves as a statistic to assess the ability of a model to distinguish between two classes. The effectiveness of the classification approach increases with the area under the ROC curve.

We compared our categorization findings to those reported in the literature, as shown in Tables 3 and 4. As converted MCI suggested AD and non-converter MCI showed stable MCI, we also compared the findings for sMCI vs. AD with those for ncMCI vs. cMCI. Most published literature regarding binary classification tasks utilized accuracy, sensitivity (SEN), specificity (SPE), balanced accuracy (BA), and AUC to demonstrate their findings.

We evaluated accuracy and AUC during our experiments. By using a confusion matrix of the best fold of the testing, we also computed SEN, SPE, and BA through the following formulas: TP/ (TP + FN), TN/ (TN + FP), and (SEN + SPE)/2, respectively.

The issue of whether patients with MCI can accurately self-diagnose their risk of developing AD remains essential to the development of viable treatments for the disease. Categorizing AD and sMCI is more challenging due to the subtler morphological changes that must be noticed, as demonstrated by the fact that the accuracy of several of the study results included in Table 2 barely reached 70–80%.

The best Level 1 learning model, reported by Suk et al. [60], had a maximum accuracy of 74.82%. It is crucial to highlight that Level 1 approaches alter spatial localization in the feature extraction process of brain imaging data, as they rely on manual feature extraction. Without taking spatial relationships into account, it is hard to guess how the model decides how to classify something in a reliable way. The Level 2 model proposed by Pan et al. [59] showed a maximum accuracy of 83.81%. They suggested a multi-view separable pyramid network (MiSePyNet), a 2D CNN model that utilizes 18F-FDG PET images. MiSePyNet was built on the concept of quantized convolution and used independent slice and spatial-wise CNNs for each view. However, this Level 2 research only used a small part of the original datasets, thus disposing of any obvious outliers and making it hard to fairly compare its performance. In another study [42] carried out by Basaia et al., 86.30% accuracy was obtained using a 3D CNN. MRI scans were segmented to create GM, WM, and CSF tissue probability maps in the MNI space. It was also built on a ROI-focused strategy rather than E2EL. Other studies [23, 27, 60] that used 2D TL with a pretrained network or local TL by transferring the knowledge of the AD vs. CN task to predict early diagnosis of AD obtained accuracies up to 82%.Only two research articles regarding multiclass categorization tasks could be found.

One by Wu et al. [12] utilized 2D MRI slices and the pre-trained 2D CNN networks CaffeNet and GoogleNet, obtaining an average accuracy of 87.00% and 83.20%, respectively. However, their implementation was based on Level 2 learning, and only obtained a 72.04% (for CaffeNet) and a 67.03% (for GoogleNet) accuracy rate for the classification of sMCI cases. Using Level 3 E2EL,MRI images as input and a basic 3D CNN model, Tufali et al. [61] conducted experiments for multiclass classification, but only obtained an average accuracy of 64.33% and an MCI class accuracy of 51.25%.

We achieved an accuracy of 93.10%in the evaluation of unseen data for the binary classification task and 87.38%for the multiclass classification task. This is significantly better than the early AD prediction accuracy reported by state-of-the-art methods in the last five years. Although our models are suitable to use in clinical settings to aid neuroradiologists, further training with more high-quality MRI scans from a diverse range of sources is required to ensure reproducibility.

Several conclusions can be extracted from the research presented here. Even with neuroimaging, where a limited number of high-dimensional scans are available, the fusion of E2EL with TL allows the obtaining of remarkable results. However, it requires the fine tuning of hyperparameters, and an appropriate 3D CNN architecture specifically designed for TL with excellent potential for generalization; additionally, MRI scans must be thoroughly pre-processed to maintain the spatial link and enhance image quality.

We also observed that MONAI offers a leading framework to implement DL models for medical image analysis, as it is simple to understand and supports a wide range of functions. Additionally, Google Colab Pro + offered the best online cloud-based resources, with access to large RAM and excellent GPUs, thus enabling us to achieve this task despite certain drawbacks such as GPU unavailability under heavy load.

The results obtained in our experiments utilizing the ADNI and IXI datasets demonstrated that our model is more effective and efficient than the current state-of-the-art models for both binary and multiclass tasks. However, there are a number of limitations to this study that need to be addressed in follow-up research.

Furthermore, the model must be implemented in clinical settings where it can be subjected to qualitative examination to determine its robustness. The number of subjects utilized to foster E2EL was still quite low. We anticipate that, as more diverse datasets become available in the future, this approach will lead to more generic learning models.