Enhancing Classification of Alzheimer’s Disease using Spatial Attention Mechanism

2.1. Datasets

There was a successful attempt in the utilization of a comprehensive dataset, representing and encompassing images representative of Alzheimer's disease at varying stages. As mentioned in the introduction, the Magnetic Resonance Imaging (MRI) scans were indicative of the general stages associated with Alzheimer’s dementia. These stages can collectively be classified into a general or wide scope under either Cognitively Normal (CN or Normal or Non-Demented) or Alzheimer’s Dementia (AD or Demented or Cognitively Abnormal).

Most research in the field of Alzheimer's disease has heavily relied on two prominent datasets: the OASIS dataset and the ADNI dataset, both of which have significantly contributed to our understanding of the disease.

The OASIS (Open Access Series of Imaging Studies) dataset comprises multiple sub-datasets, with OASIS-1, OASIS-2, and OASIS-3 being pivotal components. OASIS-1 primarily includes structural brain MRI data, serving as a valuable resource for anatomical analyses [25] and volumetric measurements. OASIS-2 extends this legacy by adding in-depth clinical and demographic information to the imaging data, facilitating comprehensive investi- gations into the progression of Alzheimer's disease. This component essentially provides longitudinal data of subjects over time, documenting the disease’s progression for future studies and early prediction of the disease. OASIS-3 is one of the more recent projects that further bolsters the collection with a broader and more diverse dataset. Besides documenting symptoms over time, it also deals with multi-modal imaging, containing a variety of imaging sessions, namely Magnetic Resonance (MR) sessions, Positron Emission Tomography (PET) sessions, and Computed Tomography (CT) sessions for patients.

In our work, we acquired the “Alzheimer's Dataset (4 classes of Images)” from Kaggle, which comprises a total of 6,400 images. These images are meticulously categorized into four distinct classes, each corresponding to different stages of Alzheimer's disease, namely: Very Mild Demented, Mild Demented, Moderate Demented, and Non-Demented. Upon further inspection, it can be noted that the dataset contains samples of data from OASIS 1 and OASIS 2, owing to it being T1-weighted MRI data of the Axial plane, a characteristic most prominently showcased by OASIS data, and containing the relevant slice of the 3D image that is also included in OASIS’s repositories as well.

A view of a sample of data from each class has been highlighted in Fig. (1).

2.2. Ethical Considerations of Proposed Research

There are certain considerations that need to be kept in mind while engaging in analysis and research involving usage of medical data repositories. These are paramount when using human subject data in research, even if the data is publicly available and de-identified. Despite efforts to anonymize data, there's still a risk of re-identification or revealing sensitive information through data aggregation. It's crucial to ensure that the data was collected ethically, with informed consent and adherence to privacy regulations.

In our research, we ensured that the dataset, although publicly available, was obtained through ethically sound practices, including proper informed consent procedures and adherence to privacy regulations. We also conducted a thorough examination of the anonymization protocols in place to safeguard the identities of the individuals whose data form the basis of our study.

Moreover, throughout our research, we've prioritized transparency and accountability. We've openly acknow- ledged the origin of our dataset, recognizing the individuals who contributed to it, and expressing gratitude for their invaluable contribution to scientific advancement. Additionally, we've implemented robust privacy measures to prevent any potential misuse or unauthorized access to the data.

Furthermore, our study design incorporates ethical considerations by focusing on the potential societal benefits of our research outcomes. We've strived to ensure that our findings serve the greater good, contributing to the understanding and diagnosis of Alzheimer's disease, while simultaneously minimizing any potential risks or harms to the individuals involved.

Continuously mindful of the ethical implications of our work, we've engaged in ongoing ethical oversight, seeking guidance from institutional review boards and ethics committees to ensure that our research aligns with the principles of beneficence, non-maleficence, and respect for persons. By upholding these ethical standards, we aim to conduct responsible research that not only advances scientific knowledge but also respects the dignity and rights of the individuals whose data we utilize.

3.1. Experimental Setup

The study used 6,400 preprocessed T1-weighted MRI pictures classified into Very Mild Demented, Mild Demented, Moderate Demented, and Non-Demented classes. The dataset was obtained from Kaggle, where the dataset incorporated samples from OASIS 1 and OASIS 2, presenting axial slices of 3D images prominent in OASIS data. No means of preprocessing or data augmentation techniques were employed besides model specific rescaling, this is due to the dataset already being preprocessed to an extent.

The Convolutional Neural Network (CNN) employed in this study featured rescaling, Conv2D layers with increasing filters up to 64, and Spatial Attention layers for relevant region focus. While Dropout layers reduced overfitting by randomly deactivating 20% of neurons, Max-pooling layers shrunk feature maps. Dense layers made it easier to capture complex patterns for image categorization, while a flattened layer came before dense layers for compatibility. The model exhibited robust feature extraction and decision-making capabilities through its usage of intricate and personalized feature mapping and specialized activation functions.

3.2. Dataset Splitting

The dataset is further partitioned into training, testing and validation subsets to facilitate the training and evaluation of any proposed model but further test our model, with this segregation being done in a specified seed.

The image ratio in accordance with the classes could be specified with 3200 Cognitively Normal scans and the other half of 6400 having very mild to moderate dementia symptoms. This brought the distribution of the images per class, as mentioned in Table 2.

3.2.1. Train Test Validation Split

The obtained Dataset was split into training, validation and testing subsets, with their respective ratio of splits being 0.8, 0.1 and 0.1, and described in Table 3. Fig. (2) illustrates the difference and effect of their splitting ratios.

Table 3 presents a breakdown of image counts across different classes in a dataset used for the classification of Alzheimer's disease. The dataset is divided into three subsets: training, validation, and test. Each subset is further categorized into four classes representing different stages of dementia: Moderate Demented, Mild Demented, Very Mild Demented, and Non-Demented. The training set comprises 51 images of Moderate Demented, 716 images of Mild Demented, 1792 images of Very Mild Demented, and 2560 images of Non-Demented cases. The validation set includes 6, 89, 224, and 320 images for each respective class, while the test set consists of 7, 91, 224, and 320 images.

3.3. Data Preprocessing

Before undergoing model development, the dataset undergoes several preprocessing steps to ensure its viability for training. Some of the images associated with the datasets are already preprocessed (i.e. the Kaggle dataset comprising 6400 images is all 128 x 128 in dimensions with justifiable 3 channel dimensions for grayscale images). Image normalization is applied to adjust the values of image pixel intensity. It scales each pixel's intensity by a factor of 1/255, ensuring that the pixel values are within the range [0, 1].

In order to improve the quality of the images, we implemented certain image processing techniques, such as thresholding but they led to a decrease in our model’s performance. Other techniques like filters and noise reduction are not suitable for medical imaging because of their predilection towards reducing essential details provided in the MRI scans that get removed and otherwise would have contributed to the performance of the CNN models.

Since the amount of data for medical imaging, especially for AD classification, is limited, various research projects have experimented with incorporating additional data augmentation techniques, which can especially benefit the distribution of images in the moderate demented class containing a lesser amount of samples as opposed to the other classes. Augmentation techniques included manipulating brightness, zoom, and rotation range, alongside flipping the image to create newer copies. But, we ultimately found that they did not enhance our proposed model and, in fact, caused a decline in our results [26].

3.4. Model Training

After performing these crucial preprocessing steps, the training set is used to train the model, the validation set for hyperparameter tuning, and the test set for model evaluation. The steps of proposed model training and classification are illustrated in Fig. (3).

3.4.2. Convolutional Neural Network

The CNN model comprises multiple layers, including convolutional 2D, MaxPooling, Dropout, and dense layers, with the SpatialAttention layer integrated at key points to focus on relevant image regions. The illustration of the layers of the CNN Sequential Model is given in Fig. (4).

Table 4.

Layer	Type	Output Shape	Parameters	Explanation
Rescaling	Rescaling	(None, 128, 128, 3)	0	This layer scales input images by a factor of 1/255 to normalize pixel values between 0 and 1.
conv2d	2D Convolution Layer	(None, 128, 128, 16)	448	Convolutional layer with 16 filters and a 3x3 kernel. Applies spatial filters to extract features from the input images.
spatial_attention	Spatial Attention	(None, 128, 128, 16)	0	Spatial attention mechanism to weight different regions of the feature map adaptively.
max_pooling2d	Max Pooling 2D	(None, 64, 64, 16)	0	Max-pooling layer with a 2x2 pool size, reducing the spatial dimensions by half, and retaining the most important features.
conv2d_1	2D Convolution Layer	(None, 64, 64, 32)	4,640	Convolutional layer with 32 filters and a 3x3 kernel. Further feature extraction and transformation.
spatial_attention_1	Spatial Attention	(None, 64, 64, 32)	0	Spatial attention applied to the features extracted by the previous convolutional layer.
max_pooling2d_1	Max Pooling 2D	(None, 32, 32, 32)	0	Max-pooling to downsample the feature map again.
dropout	Dropout	(None, 32, 32, 32)	0	Dropout layer for regularization, randomly setting a fraction of input units to zero during training.
conv2d_2	2D Convolution Layer	(None, 32, 32, 64)	18,496	Another convolutional layer with 64 filters and a 3x3 kernel, increasing the depth of feature representation.
spatial_attention_2	Spatial Attention	(None, 32, 32, 64)	0	Spatial attention applied to the features from the previous convolutional layer.
max_pooling2d_2	Max Pooling 2D	(None, 16, 16, 64)	0	Max-pooling again to further reduce spatial dimensions and capture high-level features.
dropout_1	Dropout	(None, 16, 16, 64)	0	Dropout layer for regularization to prevent overfitting.
flatten	Flatten	(None, 16,384)	0	Flattens the feature map into a 1D vector for input to the fully connected layers.
dense	Dense	(None, 128)	2,097,280	Fully connected layer with 128 neurons, introducing non-linearity and learning high-level representations.
dense_1	Dense	(None, 64)	8,256	Another fully connected layer with 64 neurons, further reducing the dimensionality of features.
dense_2	Dense	(None, 4)	260	Final fully connected layer with 4 neurons, representing the output classes for Alzheimer's classification.

Table 5.

Sr. No	Name	Type	Shape
0	rescaling	Rescaling	(None, 128, 128, 3)
1	conv2d	Conv2D	(None, 128, 128, 16)
2	spatial_attention	SpatialAttention	(None, 128, 128, 16)
3	max_pooling2d	MaxPooling2D	(None, 64, 64, 16)
4	conv2d_1	Conv2D	(None, 64, 64, 32)
5	spatial_attention_1	SpatialAttention	(None, 64, 64, 32)
6	max_pooling2d_1	MaxPooling2D	(None, 32, 32, 32)
7	dropout	Dropout	(None, 32, 32, 32)
8	conv2d_2	Conv2D	(None, 32, 32, 64)
9	spatial_attention_2	SpatialAttention	(None, 32, 32, 64)
10	max_pooling2d_2	MaxPooling2D	(None, 16, 16, 64)
11	dropout_1	Dropout	(None, 16, 16, 64)
12	flatten	Flatten	(None, 16384)
13	dense	Dense	(None, 128)
14	dense_1	Dense	(None, 64)
15	dense_2	Dense	(None, 4)

The initial layer involves rescaling the inputs, converting and normalizing the pixel values to a range of [0,1]. The Conv2D Layers involve a sequence of convolutional layers, extracting features from the input images. Initial layers have 16 filters, slowly doubling in every layer to a maximum of 64, which corresponds to the grid size for every layer. After each convolutional layer, the Spatial Attention layer is applied to highlight significant regions of the feature maps (Eq 1).

(1)

µ_c is the mean of channel c.

H is the height of the feature map.

W is the width of the feature map.

x_c(i, j) is the value of channel c at position (i, j).

Sigmoid function calculates (Eq 2).

(2)

Mathematically, this line performs element-wise multiplication between the input tensor and the attention tensor for each channel c (Eq 3):

(3)

This is returned back to the model.

Where:

y_c(i, j) is the output value for channel c at position (i, j).

x_c(i, j) is the input value for channel c at position (i, j).

a_c is the attention score for channel c, which scales the input value.

Max-pooling layers reduce the spatial dimensions of the feature maps while preserving important features. In a max pooling (2,2) layer, the feature map is divided into individual regions or windows of size (2,2) pixels. These windows slide over the feature map, moving by 2 pixels horizontally and 2 pixels vertically at a time (in this case). Within each window, the operation selects the maximum value. This maximum value represents the most important feature in that specific window. The selected maximum values from each window are then used to create a new feature map with reduced spatial dimensions. These maximum values are then compiled to construct a new feature map that is half the size of the original one, both horizontally and vertically.

Dropout layers are employed after every MaxPooling layer to prevent overfitting by randomly deactivating a fraction of neurons during training, which can be specified in the model. The optimal value selected for our model was 0.2, which is also a common value taken in other image classification projects. A dropout signifies that for each training example and at each training step, dropout randomly “drops out” (or deactivates) 20% of the neurons in the specified layer. In other words, 20% of the neurons in that layer are set to zero during that training step. This has an effect similar to ensemble learning.

One of the penultimate layers involves a flattened layer, which is applied after the last dropout layer and before the Dense Layers. Flattening in deep learning converts multidimensional data, like the vectors of images or feature maps, into a one-dimensional vector. It's crucial for making this data compatible with fully connected layers and creating a feature vector for tasks like classification. It also decreases the dimensionality and size of the original input vector it processes.

In the context of classifying images, dense layers are used to capture intricate patterns and features extracted from previous layers, enabling the neural network to understand and make decisions about the content of the images. They connect every neuron in a layer to every neuron in the subsequent layer, creating a complex web of connections. This connectivity between the neurons allows them to learn and model intricate relationships within the image data due to the extensive connections. Each neuron in a dense layer computes a weighted sum of inputs and applies an activation function (commonly ReLU), making them capable of capturing various features and patterns in the data.

While the first Dense layer contains 128 neurons in its layers, the number progressively decreases to 64 in the second layer and a final 4 neurons in the final layer. These 4 neurons mimic the number of classes for classification for which a softmax activation function is used.

These layers, as illustrated in Fig. (5), serve as the final step in the network, using the features to classify the images into different categories, making them crucial for image classification tasks .

4.1. Performance Evaluation of Proposed Architec ture

During training, the model achieved a remarkable accuracy of 0.9965 ± 0.003. By the end of training, the model's loss had reached an impressive minimum of 0.0098, reflecting its successful optimization. Notably, it seemed to flatline around the 60th epoch, indicating that the model had achieved stable convergence.

Moving on to the validation set, the model maintained an impressive accuracy of 0.9953 ± 0.0016. The validation data loss, reflecting the generalization capability of the model, remained low at 0.0095 ± 0.004.

Our proposed model has proven to be remarkably accurate, demonstrating its performance in capturing patterns within the training data and its impressive ability to generalize to new data. These results serve as a further testament to the robustness of our model, showcasing its reliability and impressive performance on the validation set. Fig. (6a-c) represents the accuracy and loss dynamics of both the training and validation datasets, which can help enhance our understanding of the proposed model's performance.

Fig. (6) provides the accuracy and loss. Fig. (6a) showcases the comparison between the increasing accuracy and decreasing loss of the training dataset over successive epochs on the x-axis. Notably, the loss curve steadily diminishes, approaching nearly zero (0.0098), while the accuracy curve ascends to a near-perfect value, stabilizing at around 0.9965.

Fig. (6b) shows a comparison between the accuracy curves of training and validation data, offering insights into how well the model generalizes to unseen instances. Fig. (6c) focuses on the comparison between the loss curves of training and validation data.

But when it came to the testing dataset, an accuracy of 1.0 i.e. 100% was obtained, with a loss on the test dataset equal to 0.0051. The confusion matrices are shown in Fig. (7a and b) on testing data and validation data. The number of misclassifications is closer to zero in all the classes in the case of testing data. However, in the case of validation data, there is one misclassification in the case of non-demented and very_mild_demented classes.

4.1.1. ROC AUC Scores

The Receiver Operating Characteristic (ROC) curves for each class shown in Fig. (8a) reflect a highly effective classifier on the validation test. With Area Under Curve (AUC) values close to 1.0, the model demonstrates robust discriminatory power in distinguishing between different levels of dementia and non-demented cases. This performance underscores the model's reliability and accuracy in the classification task for dementia severity. The ROC curve for the testing set shown in Fig. (8b) reveals a model that performs at the highest level of accuracy for each class, achieving perfect discrimination between different levels of dementia severity and non-demented cases. This outcome underscores the model's exceptional reliability and effectiveness in dementia classification on the testing data.

4.2. Comparison with State-of-the-art Techniques

Our research paper provides a thorough evaluation of different convolutional neural network (CNN) structures for classifying Alzheimer's scans into four specific categories: Very Mild Demented, Mild Demented, Moderate Demented, and Non-Demented. The results of our experiment present essential performance metrics such as test accuracy, sensitivity, and specificity, which offer crucial insights into the effectiveness of each model.

4.3. Result Reproducibility and Transparency

The aforementioned results were obtained through the usage of specific dataset splits for training, testing and validation, which involved breaking down the dataset into a training set, a testing set as well as a validation set, in the ratio 0.8: 0.1: 0.1 as visualized in Fig. (2). The segmentation and distribution of images into a particular set was done so through python’s train test split module. One of the parameters in the function involves a ‘random_state’ variable which, when set to a specific number, ensures reproducibility of the ratio and distribution of these images. With means to the Convolutional Neural Network model, the hyperparameters were set based on the optimal performance of the model. These hyperparameters include filters, kernel_size, padding, and activation functions, all of which can be reproduced based on the specific model setup required.

5.1. Unique Contributions and Practical Implications

Our study aims to improve the precision of identifying Alzheimer's disease by incorporating a spatial attention mechanism into a sequential CNN model. This stands apart from conventional models that lack this critical element. Our innovative methodology has shown significant success, surpassing the effectiveness of traditional models in key metrics such as test accuracy, sensitivity, and specificity – achieving a perfect test accuracy, as well as sensitivity and specificity of 100%. Our proposed algorithm boasts an impressive novel feature: the inclusion of a spatial attention mechanism. This proved to be a highly effective and promising strategy for enhancing the classification of Alzheimer's disease. Our study not only evaluates well-established CNN architectures but also represents a testament to an innovative and intuitive approach. By incorporating spatial attention mechanisms, we have made significant progress in the realm of Alzheimer's diagnosis through image classification. The impressive results of our algorithm serve as a reminder of the crucial role spatial relationships play in medical image analysis [25], especially in the context of neurodegenerative diseases.

The practical implications of our research are significant. Our algorithm has demonstrated remarkable accuracy, sensitivity, and specificity, indicating its potential for practical use in clinical settings. By utilizing this enhanced performance, medical professionals can greatly benefit in accurately and efficiently diagnosing Alzheimer's disease through neuroimaging data [27]. Furthermore, the incorporation of spatial attention mechanisms presents exciting opportunities for future advancements and exploration in the field of medical image analysis.

In conclusion, our study not only evaluates established CNN architectures, but also introduces an intuitive approach that surpasses the performance of traditional models. The addition of spatial attention mechanisms in our algorithm makes a notable contribution to the field of Alzheimer's diagnosis through image classification.

5.2. Limitations

The spatial attention layer operates by computing a set of operations on the average pixel values of each channel, following which it undergoes several other transfor- mations through functionsun through a sigmoid function, resulting in a spatial attention map. This spatial attention map is a matrix of weights that are used to regulate the values of the input feature map. The resulting attended feature map is then used by the subsequent layers of the CNN.

This set of operations remains uniform for every passed input. So far, the utilized data comprises 6400 test subjects to which our model has generalized adequately, taking into account specific regions which may hold higher propensity towards Alzheimer’s stage detection.

This can help only when being used on the same form of imaging as utilized in our model training, i.e. T1 weighted Brain MRI of the axial plane. However, in order to achieve our future scope of being able to map out progressive classification over time for Alzheimer’s patients, the model would be required to be trained using MRI training images for patients over time, for which the current model is not able to generalize available data.