Towards practical application of deep learning in diagnosis of Alzheimer's disease

Abstract

Background

Accurate diagnosis of Alzheimer's disease (AD) is both challenging and time consuming. With a systematic approach to diagnosis, steps can be taken toward improved treatment and prevention of the disease.

Objective

This study explores the practical application of deep learning models for the diagnosis of AD across different disease stages.

Methods

Due to computational complexity, long training times, and limited availability of labeled datasets, full brain three-dimensional (3D) convolutional neural networks (CNNs) are not commonly used, and many studies rely on two-dimensional (2D) variants. In this work, full brain 3D versions of well-known 2D CNN architectures were designed, trained, and tested for the diagnosis of multiple stages of AD. More than 1500 full brain volumes were used for model training and evaluation.

Results

The proposed deep learning approach demonstrated good performance in differentiating various stages of AD. In addition to classification, the models were able to extract discriminative features relevant to disease stage. These features aligned with meaningful anatomical landmarks that are currently considered important for AD identification by clinical experts. An ensemble of all algorithms was also evaluated and achieved superior performance compared to individual models, with a maximum classification accuracy of 87.4%.

Conclusions

The trained 3D CNNs and their ensemble show strong potential for assisting in the diagnosis of AD. These models may be incorporated into clinical software tools to support physicians and radiologists in improved diagnostic decision making.

Introduction

Deep learning is a type of artificial intelligence algorithm which is capable of extracting useful features from the data without the use of dedicated feature extraction tools. It has found applications in areas such as self-driving cars, natural language processing, computer vision, etc. ¹ This study focuses on practical application of deep learning techniques for early detection of Alzheimer's disease (AD). In this study, three-dimensional (3D) deep learning algorithms are designed and trained for direct application to 3D neuroimaging data. These 3D algorithms were obtained by modifying the two-dimensional (2D) counterparts which are commonly used for images.

Human brains have 3D structures. Using a 2D model for analysis of brain structures may cause some loss of information that extends across slices. Using 3D models preserves special information but practical application of 3D deep learning models is challenging due to availability of limited dataset, increased computational complexity and increased training time. Various approaches have been used to indirectly incorporate 3D information with deep learning. Here we propose the use of a direct 3D deep learning approach for feature extraction and classification of various stages of AD using minimally preprocessed full brain 3D anatomical magnetic resonance imaging (MRI) data.

Over the last decade, deep learning has been increasingly used in neuroimaging with more than 2100 studies (https://pubmed.ncbi.nlm.nih.gov/?term=%28alzheimer%27s+disease%29+AND+%28deep+learning%29&filter=years.2016-2025) applying deep learning for AD classification in last decade. A complete comprehensive review of all the deep learning literature with application to AD is out of the scope for this article. However, some of the different deep learning approaches for AD are discussed in detail in the Supplemental Material.

Convolutional neural networks (CNN) are among the state-of-the-art machine learning algorithm for classification, segmentation and generation of images.^2,3 However, the majority of the CNN development is centered around images (2D data). To make use of CNNs for volumetric data, especially medical imaging scans, modifications have to be made to the data and/or the CNN to make them compatible with one another.

Some of the early works for classification of AD used 2D CNN on anatomical images in the axial, ⁴ sagittal, ⁵ or coronal ⁶ planes. Some studies used 2D slices from the entire brain while others used slices overlapping in the hippocampus region ⁷ or based on entropy information. ⁸ When using only 2D images, spatial information is lost. Some early researchers tried to overcome this limitation by using simpler 3D CNN with single region volume, like hippocampus, ⁹ or 3D patches of smaller size. ¹⁰ Apart from MRI, deep learning models have also been used with other modalities like positron emission tomography (PET) ¹¹ and magnetoencephalography (MEG). ¹² Deep learning models have also been used with 4D functional MRI (fMRI) data to classify different stages of AD. Multichannel 3D CNNs have been used for both binary ¹³ and multiclass ¹⁴ classification of AD using fMRI data.

Different models take advantage of different aspects of neuroimaging data. Broadly speaking there is an indirect trade off between network training complexity and data preprocessing complexity. As major portion in training a CNN is a one-time aspect, while prediction on new datapoints is frequent and thus needs to be fast. Hence, we believe that it is more advantageous to have a network which takes more time to train but is relatively simpler and faster for the actual application purpose.

We designed and trained 3D versions of three well known CNN architectures and compare their performance for neuroimaging data obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The three CNN architectures—Alexnet, ¹⁵ GoogleNet, ¹⁶ and VGG 16 ¹⁷ —were choosen for their distinct characteristics and popularity. Alexnet is the simplest CNN architecture with series of convolutional and pooling layers for feature extraction followed by fully connected layers for classification. The unique aspect of GoogleNet architecture is the Inception module, consisting of parallel convolutional layers. The VGG16 architecture consists of cascaded (series) convolutional layers between pooling layers.

Along with comparison of performance of all three types of CNN (simple, parallel, and series convolutions), the feature extraction of all CNNs were investigated. Features that each network considered significant for making a prediction were studied to find similarities with known anatomical features associated with AD. Finally, with practical application in mind, an ensemble framework is presented which improves the diagnostic accuracy for classification of AD. The presented framework has a potential to be developed into a stand-alone software package that can provide additional information along with the traditional brain images.

Methods

Dataset and preprocessing

The CNN were trained using anatomical MRI data. Data used in the preparation of this article were obtained from the ADNI database (adni.loni.usc.edu). The 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 MRI, PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early AD. Anatomical data for 1502 participants were downloaded from the ADNI online database. The participants were from 3 different groups: Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), and healthy control (CN). The group wise demographic data is shown below.

AD: 225 F / 275 M, age: 75.74 ± 7.82 years

MCI: 217 F / 288 M, age: 75.54 ± 8.12 years

CN: 213 F / 284 M, age: 75.33 ± 7.31 years

All the anatomical data were acquired on a 3T MRI machine with MPRAGE, 3D MPRAGE or accelerated MPRAGE imaging protocol. The online dataset consists of data from different scanning sites and thus the data dimensions are different. As preprocessing, all the volumes were normalized to a standard MNI brain atlas ¹⁸ with a spatial resolution of 1 mm × 1 mm × 1 mm and segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). SPM 12 toolbox was used along with MATLAB 2020a to do the preprocessing. ¹⁹ The segmented volumes acted as 3 channel inputs to the CNNs. Thus, the input size to the 3D CNN was 157 × 189 × 156 × 3, which is equivalent to 13.8 million datapoints per brain. For every input, the network outputs the probability of that input being in one of the three categories: AD, MCI, or CN. A total of 1502 brains were used to train, validate and test the CNNs resulting into about 21 billion individual datapoints. Each of the CNN also consists of tens of millions of learnable parameters which were optimized (trained) for thousands of iterations thus making it a big data problem.

Convolutional neural network

Architecture. Some of the most commonly used CNNs are designed for 2D images while neuroimaging data is typically 3D. When using a 2D CNN directly on to sections of 3D brain important spatial information is lost which may reduce diagnostic efficiency. Thus, 3D CNNs are more suitable for neuroimaging applications. A major contribution of this study is the implementation and training of 3D CNN. In this study a 3D version of existing 2D CNN is being implemented and trained for classification of 3D full brain MRI data into different stages of AD. Three well known CNN architectures trained in this study are Alexnet, GoogleNet, and VGG16. The three CNN architectures were modified to handle 3D data while keeping the base architecture as close to the original 2D image version as possible.

Each of the 3 selected CNN have a unique characteristic in terms of data flow through different layers. In 3D variant, uniqueness of the interconnection of the layers were kept the same. However, changes were made to handle an extra dimension in the data. By simply adding an extra dimension, both time and computational complexity increases significantly. For example, a 3 × 3 kernel convolution for 64 × 64 image will require 34,596 (62 × 62 × 9) multiplications and 3844 (62 × 62) addition operations. Just by increasing from 2D to 3D, the same sized (3 × 3 × 3) kernel convolution will require 6,434,856 (62 × 62 × 62 × 27) multiplications and 238,328 (62 × 62 × 62) addition operations.

Thus, just adding an extra dimension will increase the computation time and computational complexity by orders of magnitude making the training process impractical. Hence the scaling from 2D to 3D was be done in a way to keep the computation time and complexity within practical limits of the hardware.

The most number of convolution operations (multiplications and additions) are performed in the earlier stages of the CNN when the input image size is the largest. Thus, initial layers were optimized the most for 3D counterparts to keep the computation time and complexity within practical limits. Moreover, a direct reduction in computational complexity is achieved by reducing the channel depth at each layer. Modification of the pooling window and padding size also reduced the computation time and memory requirement of the overall network.

The size of the padding and pooling was also crucial in determining the input size of the successive layers. The input size of the CNNs were also modified to make it suitable to handle 3D anatomical brain data from any MRI scanner. A common preprocessing step in neuroimaging analysis involving multiple participants is normalization to a global atlas. Normalization allows data from different sites and participants to be in a uniform spatial dimension. The input size of the CNNs were thus selected to be the size of the standard MNI atlas which will allow the networks to be used on practically any normalized anatomical MRI data.

Another common feature of the 2D CNN used for images is the color channels for the input layers. Color images consist of red, green, and blue channels but neuroimaging data only consists of intensity information. However, brain volumes can be decomposed into various tissue types by the process of segmentation. Segmentation is yet another common preprocessing technique used with neuroimaging data. The three main tissue types in the brain are GM, WM, and CSF. Each of the three have unique properties and thus the segmentation maps corresponding to GM, WM, and CSF are used as channels in the 3D CNN. AD is a neurodegenerative disease where there is a loss of neuron structure which in turn causes direct changes in the GM and CSF volumes.^20–22 Thus, having GM, WM, and CSF volumes as channels may result in the CNN learning clinically meaningful features.

The memory requirements were also optimized for the 3D CNN. The memory required to store a CNN depends on the total number of parameters. The total number of learnable parameters for 2D AlexNet, GoogleNet, and VGG-16 is 61 M (million), 7 M, and 138 M respectively while the 3D version has 16.8 M, 11.1 M, and 46.2 M respectively. The memory requirement for 3D Alexnet, GoogleNet, and VGG16 is 64.27 MB, 42.74 MB, and 176.4 MB respectively while the standard 2D counterparts have a memory requirement of 227 MB, 27 MB, and 515 MB respectively. If the CNN was simply scaled up from 2D to 3D, the total number of learning parameters and the memory requirements would have increased significantly. However, due to reduction in number of channels and modifications in kernel size, pooling window and padding size, the total number of learnable parameters in the 3D CNN were less than (AlexNet and VGG16) or comparable (GoogleNet) in number to their 2D counterparts. Thus, the memory requirement for storing the 3D CNN would be similar to that of a standard 2D CNN.

The architecture for all 3 CNNs is shown in Figures 1 and 2. Figure 1 shows the architecture for 3D Alexnet and 3D VGG 16 CNN, while Figure 2 shows the 3D GoogleNet architecture along with the details of the Inception subnetwork. Different colors indicate different blocks of the CNN while the arrows indicate the dataflow direction. The size of the convolutional filters is denoted for each region with the number of channels indicated within the parenthesis.

Figure 1.

(a) Architecture of the 3D AlexNet CNN. (b) Color legends for different layers depicted in the figure. (c) Architecture of the 3D VGG16 CNN. [The size of the convolutional filters and the number of filters (in parenthesis) for each of the convolution layer is denoted within each layer. The number of nodes for all the fully connected layers is denoted within the layer except for the last layer which has 3 nodes corresponding to the 3 class. All the max pooling layers were of size 3 × 3 × 3 with a stride of 2 × 2 × 2.]

Figure 2.

(a) Architecture of the 3D GoogleNet CNN. (b), (c), (d) Architecture of the Inception module at Layer 3, 4 and 5 respectively. (e) Color legends for different layers depicted in the figure. [The size of the convolutional filters and the number of filters (in parenthesis) for each of the convolution layer is denoted within each layer. The number of nodes for the fully connected layers is 3 corresponding to the 3 class. The max pooling in the GoogleNet layers were of size 3 × 3 × 3 with a stride of 2 × 2 × 2, while within the inception layers was of size 3 × 3 × 3 with stride of 1 × 1 × 1.]

Alexnet is one of the earliest and one of the simplest CNN. Like the original architecture, the 3D version has 5 convolutional and 3 fully connected layers. The VGG16 architecture is ‘deeper’ than Alexnet. The 3D VGG16 architecture has 16 learnable layers just like the original one. The uniqueness of VGG architecture is the blocks of cascaded convolutional layers. The uniqueness of GoogleNet is the Inception subnetwork. The inception subnetwork has parallel learnable layers which extracts features at different scales. Unlike the 2D GoogleNet, output is only computed at the end of the last learnable layer and not after every inception subnetwork.

Training. The entire dataset was randomized and divided into three parts: training (70%: 1052 samples), validation (15%: 225 samples), and testing (15%: 225 samples). The same split was used for all the three CNNs to be able to have a direct comparison of the performance once trained. All three CNNs were trained using MATLAB on the university's high performance computing center facility. All CNNs were trained using the ‘adam’ optimizer. ²³ The networks were trained for a total of 1024 epochs. The initial learning rate was set to 1 × 10⁻⁵ (1e-5), with learning rate reducing by a factor of 0.75 after every 256 epochs. The training was performed in minibatch of size 32 with all training samples being randomized after every epoch. The validation frequency was set to 128 iterations which is roughly equivalent to 4 epochs. To avoid overfitting, L2 regularization was used with a regularization constant (λ) of 0.1. Overfitting was also reduced by using dropout layers after every fully connected layer for all 3 CNNs. The dropout rate was set to 0.5. A 5-fold cross validation was also performed for each of the three CNN to check if the network could generalize well. The training time taken for the validation accuracy to reach at least 80% was 14 h for Alexnet, 20 h for GoogleNet, and 35 h for VGG16.

Analysis. Once trained, performance parameters were computed for all three CNNs using the confusion matrix (for validation and testing partitions). The confusion matrices were used to calculate the class-wise accuracy, precision, sensitivity, specificity, and F1 score. Class-wise ROC (receiver operating characteristic) curves were also generated, and corresponding AUC (area under the curve) was calculated. The performance parameters were also computed for each of the 5 folds of the 5-fold cross validation. Because the same data split was used for training of each of the three CNNs, the predictions from the three CNNs were combined using an ensemble algorithm. The ensemble algorithms were applied to data samples not used for training, i.e. combined validation and testing split. Ensemble averaging and voting ensemble algorithms were compared. In ensemble averaging, the class-wise prediction probabilities were averaged to obtain ensemble probabilities for each class, and the input sample was assigned to the class with highest ensemble probability. In voting ensemble, the input sample was assigned to the class with the highest number of prediction votes. Because there are 3 class and 3 separate CNN models, it is possible that each class may get a single vote. In that case, the input will be assigned to the class with highest prediction probability among all three CNNs.

For each input sample, every classifier produced a probability distribution representing the likelihood of that sample belonging to each diagnostic category (AD, MCI, or CN). The predicted class corresponded to the category with the highest probability. To better understand model behavior, a misclassification analysis was performed for all samples that were incorrectly classified. For each of these samples, we recorded the true class and the predicted probabilities across all categories. The second- and third-highest probability classes were examined to identify patterns in near-correct predictions. Additionally, histograms were generated to visualize the distribution of predicted classes for each true class across the three CNN architectures. This analysis helped determine whether errors occurred between clinically adjacent categories or were random misclassifications.

Results

The summary of the classification performance for all the three CNN is shown in Figure 3. The left and the right part of the figure shows class wise ROC curve and the corresponding confusion matrix for validation and testing data split respectively. The middle portion shows the plot of class wise performance parameters for both validation and testing data split. For the confusion matrix, the horizontal axis represents the true class while the vertical axis represents predicted class. From the figure it can be observed that all the classifiers are performing well in distinguishing between different classes. The best classification performance is observed for AD class followed by CN and then MCI. The results of the 5-fold cross validation are shown the Supplementary File. The overall high values of accuracies indicates that each of the CNN architecture is capable of performing the classification task in a satisfactory manner.

Figure 3.

Performance parameters of validation and testing data split for all 3 CNN architecture (rows). The plots on the left and right are the class-wise ROC curve for validation and testing data split. The confusion matrix is also shown on the subplot The colors correspond to the value in the cell. Larger numbers are given lighter color (yellow) while smaller numbers are given darker color (blue). The bar chart in the center shows class-wise network performance parameters (Acc: accuracy; Pre: precision; Sen: sensitivity; Spc: specificity; F1: F1-score; AUC: area under the curve).

The performance of the ensemble algorithm is shown in Figure 4. In the figure, performance parameters of both the ensemble approach are compared with individual CNN architectures. As expected, the ensemble approach performs better than individual CNN. Moreover, the performance of both the ensemble approach is almost similar. Table 1 shows the performance comparison of a few 3D CNN techniques for classification of AD. All the studies shown in Table 1 uses 3D CNN with either full or a section of MRI/fMRI brain volumes. Some of the studies uses the CNN as feature extractor and then uses sophisticated techniques like support vector machine (SVM) or long short term memory (LSTM) to make the classification prediction. It can be observed from the table that in general binary classifiers performs better than multiclass classifiers. The performance of the 3D CNNs discussed in the current article is comparable or slightly better than other studies using 3D CNN with full brain MRI.

Figure 4.

Performance comparison of individual CNNs and ensemble algorithm. The x-axis represents true class while the y-axis represents parameter value.

Table 1.

Comparison of a few 3D CNN Alzheimer's disease classifiers based on input modality and output type. AD: Alzheimer's disease; CN: healthy control; MCI: mild cognitive impairment; pMCI: progressive MCI; sMCI: stable MCI; MRI: magnetic resonance imaging; fMRI: functional MRI; LSTM: long short term memory; SVM: support vector machine.

Study	Modality	Type	# Participants	Accuracy
Liu et al. 9	3D MRI patch	Binary	449	80.1% (AD vs. CN)
				71.5% (MCI vs. CN)
Oh et al. 24	3D MRI	Binary	695	85.2% (AD vs. CN)
				77.3% (pMCI vs. CN)
				63.1% (sMCI vs. CN)
				60.9% (pMCI vs. AD)
				75.1% (sMCI vs. AD)
Raju et al. 10	27 3D MRI patches + SVM	Multiclass	465	97.70%
Hosseini-Asl et al. 25	3D MRI	Multiclass	210	89.10%
Li et al. 3	3D fMRI	Multiclass	389	82.46%
	3D fMRI + LSTM			89.47%
Current Study	3D MRI	Multiclass	1502	84.3% (3D Alexnet)
				83.3% (3D GoogleNet)
				84.5% (3D VGG16)
				87.4% (Ensemble)

Figure 5 shows the class-wise (columns) misclassification histograms for all three CNN architectures (rows). For true class AD, nearly all misclassified samples were predicted as MCI, while for true class CN, most misclassifications (except for GoogleNet) were also labeled as MCI. This consistent pattern indicates that the networks rarely confused AD with CN directly. As noted in the clinical literature, sharp demarcations between normal cognition, MCI, and dementia are inherently difficult, and clinical judgment is often required to make these distinctions. ²⁶ Hence, the tendency of the models to misclassify AD or CN as MCI reflects biologically meaningful uncertainty rather than random error.

Figure 5.

Class-wise histogram for the misclassified samples for all 3 CNN architecture. Each column subplot represents a different class while each row subplot represents different CNN architecture. The x-axis shows the predicted class (Pred class) while the y-axis shows the percentage misclassified for a given true class.

For the MCI group, misclassifications occurred in both directions with some labeled as AD and others as CN. This is consistent with the intermediate and progressive nature of MCI. The results suggest that the deep learning networks are capturing a continuous cognitive spectrum rather than rigid categorical boundaries. All these misclassification results suggests that the deep learning networks are actually making predictions which are sensible and not just random. Such misclassification events can be fancily called ‘task failed successfully.’

Features that were weighted more by the deep learning networks were analyzed. Saliency maps were generated for each network and features were extracted for each class.^27,28 Figure 6 shows the saliency map of important features for each of the three network and each category. One important observation is that all three networks have some overlap with the hippocampus region and consider that as an important feature for classification of AD. Hippocampus has been shown to be a key brain region of interest for detection and classification of AD.^29–33 All the three CNNs giving more importance to the hippocampus regions indicate that the networks are actually extracting useful features to make the decision about the category.

Figure 6.

Saliency map for (top) AlexNet, (middle) GoogleNet and (bottom) VGG16 architectures mapped onto MNI brain atlas. Maps for different categories are represented in different colors. The strength of the color depends on the strength of the output of the CNN features (arbitrary value). The color coding is additive in nature.

Saliency maps for AlexNet and VGG16 overlap largely with the right hippocampi while that of GoogleNet overlaps in both the hemispheres. The difference in the saliency maps between GoogleNet and the other two architecture may be due to the difference in the network architecture itself. Alexnet and VGG16 have a more serial dataflow with earlier layers extracting smaller and deeper layers extracting larger features. On the other hand, the ‘inception’ layers of GoogleNet are capable of extracting multiscale features providing the network to extract both smaller and larger features simultaneously.

For GoogleNet, the saliency maps also include parts of the caudate nucleus for AD and CN categories. The volumetric analysis of the caudate nucleus has shown significant difference between AD and CN categories in previous studies.^34,35 It has also been shown that the absolute caudate nucleus volume and the ratio of caudate nucleus to hippocampus volume increases for AD. ³⁶ An interesting observation is the overlap of saliency maps with the cerebellum, mainly for the MCI category. Some studies do suggest degenerative changes of the cerebellum during MCI ³⁷ and AD dementia in general. ³⁸

Again, observing meaningful anatomical features in the saliency maps indicate that the networks are doing a decent job of feature extraction and some of the features can be post processed and used by human experts to make further inferences. These maps enhance interpretability and clinical usability by showing where each network focuses its attention, allowing radiologists to re-examine subtle structural changes in the highlighted regions and transforming the model from a black-box classifier into a human–AI decision-support tool.

Discussion

The present study demonstrates that 3D CNNs trained on minimally preprocessed anatomical MRI volumes can effectively classify AD, MCI, and CN participants. The ensemble model achieved higher overall diagnostic accuracy than any single architecture, highlighting the benefit of combining networks with different feature extraction mechanisms.

The performance comparison across AlexNet, GoogleNet, and VGG16 architectures shows that each network captures unique spatial and structural characteristics of neurodegeneration. AlexNet, with its simpler and sequential architecture, demonstrated faster convergence but relatively lower sensitivity to subtle features. GoogleNet, with multi-scale inception layers, performed better in capturing distributed patterns across brain regions, while VGG16's deep cascaded structure provided finer granularity in texture and boundary representation. The ensemble approach leveraged the complementary strengths of these architectures, leading to more consistent and stable classification outcomes across folds.

The results also suggest that 3D CNNs trained on full-brain volumes are capable of extracting disease-relevant information without the need for explicit ROI selection or handcrafted features. This approach simplifies preprocessing requirements, improving practical applicability in clinical workflows. Since most clinical MRI scans are already preprocessed with tissue segmentation (GM, WM, and CSF), the proposed framework can be directly integrated into existing imaging pipelines with minimal modification. Moreover, as training is a one-time computationally intensive task, the model can generate predictions quickly during deployment, supporting efficient and scalable use in diagnostic environments.

The saliency maps generated for each model revealed meaningful activation in regions such as the hippocampus, parahippocampal gyrus, fusiform gyrus, and thalamus, areas well-documented to be affected in AD pathology.^29–33 These overlapping activations across networks suggest that the CNNs are not relying on arbitrary image artifacts but are indeed learning biologically relevant representations. Clinically, this feature-level interpretability enhances user confidence and enables integration into radiological workflows as a decision-support tool rather than a black-box classifier.

The findings of this study are consistent with previous research showing the usefulness of deep learning for AD classification. However, most earlier approaches used 2D CNNs or patch-based methods, which often lose important volumetric information. By using full 3D volumes, the current framework keeps spatial continuity and anatomical context, improving how well the model can be interpreted. Compared with earlier studies using the same ADNI dataset, the achieved accuracy and balanced performance across diagnostic groups are within or above the typical range as shown in Table 1. The pattern of misclassifications seen in this study, where AD and CN are rarely mixed and MCI lies between them, matches the clinical progression of cognitive decline rather than fixed categories. This supports the biological validity of the model's predictions.

Limitations

Despite the strong performance of our volumetric CNN framework, several limitations warrant discussion. First, our model was trained and evaluated solely on the ADNI database, which is a well curated and homogeneous cohort. As noted in prior studies, models developed on ADNI may not generalize well to other populations, scanning protocols, or clinical settings with greater heterogeneity. ³⁹ Therefore, the external validity of our findings remains to be demonstrated. To partially mitigate this limitation, we included MRI scans acquired from multiple scanner manufacturers (Siemens, GE, and Philips) within the ADNI dataset to introduce hardware diversity and minimize overfitting to device-specific signal characteristics.

Second, the relatively lower classification accuracy for the MCI category is another area which needs further attention. The structural changes in MCI are subtle and vary widely across individuals, often resembling features of both cognitively normal and AD participants. ⁴⁰ Although, a lot of misclassifications is a result of not correctly identifying features related to MCI, it is possible that some of the misclassifications may arise from label uncertainty rather than algorithmic error. Clinical diagnoses in ADNI depend on expert consensus and cognitive assessment, yet the boundaries between MCI and AD are subjective and can shift with progression. ⁴¹ Subjects near this threshold may exhibit mixed features that the model detects but that categorical labels cannot fully represent.

Finally, we also used saliency maps to interpret how the networks made their decisions, but their explanatory power is still limited. As noted by Dyrba et al., ⁴² there is no clear “ground truth” for validating these maps, and a strong overlap with known brain regions does not necessarily mean the model has captured the true biological cause of the disease. Therefore, while saliency maps help visualize which areas influenced the prediction, their clinical reliability and biological meaning should be interpreted with caution.

Future work

The results from the current study suggest that full brain anatomical MRI volumes can be used along with existing deep learning algorithms to make predictions about the stages of AD dementia. The ensemble algorithm performs better than any single CNN and the extracted features by each of the networks have some meaningful anatomical information. The future work for this study is twofold. First, train the models with a larger dataset and have a provision for constant update of the CNNs, i.e. continuous online training with new input data. The network can be trained to differentiate between more categories like early and late MCI or MCI due to AD. Second, development of a software tool that can generate predictions and output feature maps which can then be used by a radiologist/expert to assist in better decision making. Instead of a hard classification, the software can output probabilities of various categories and then the human expert can make an educated decision. For instance, for a given brain volume, the CNN can output the probability of various predicted class along with the saliency maps and then a radiologist can examine both raw and processed data an make the final decision. Because the networks only need to be trained once, the post processing/prediction with new data should be fast (in terms of computation time) and direct (in terms of computational complexity).

Conclusion

In this study, it has been successfully shown that different kinds of deep learning algorithms can be used to predict/classify various stages of AD dementia using standard anatomical MRI brain images (3D volume). Well known CNNs used for 2D images can be modified and used with 3D brain volumes to detect useful neurological information. The ability to detect intermediate stages of AD like MCI may be useful for preventing or delaying the onset of AD. Such an early detection ability may prove to be very beneficial for the patient. One of the main features of the 3D CNN discussed here is the simplicity on the user's end. Once trained, the input to these CNN is full brain volume which can be obtained directly from the MRI scanner. An inbuilt preprocessing and trained network can be then deployed as a software package for the radiologists and physicians who may not be very proficient with the nitty gritty of computer programming. The performance of the ensemble algorithm is more promising than any individual CNN and the features (saliency maps) extracted by these CNNs can provide additional information to the radiologists/physicians. A human deep learning synergy can be beneficial in diagnosis of neurodegenerative disorders like AD.

Supplemental Material

sj-docx-1-alr-10.1177_25424823261415808 - Supplemental material for Towards practical application of deep learning in diagnosis of Alzheimer's disease

Supplemental material, sj-docx-1-alr-10.1177_25424823261415808 for Towards practical application of deep learning in diagnosis of Alzheimer's disease by Harshit Parmar, Eric Walden and in Journal of Alzheimer's Disease Reports