A novel approach to recognition of Alzheimer’s and Parkinson’s diseases: random subspace ensemble classifier based on deep hybrid features with a super-resolution image

Listed below are the key contributions of the proposed model in the classification of Alzheimer’s and Parkinson’s diseases

• The current study adapted deep super-resolution neural network architecture for MRI images. Thus, high-resolution, and enhanced MR images were obtained.

• A hybrid deep neural network extracted low and high-level features from MRI images.

• We presented a high-performance KNN-based random subspace ensemble classifier model in classifying AD, PD, and healthy control.

• Experimental studies have shown that our system is a more accurate, faster, less costly, and more objective approach. Based on these findings, it seems likely that our AI system could assist in diagnosing patients in clinical settings in the future.

The remainder of the article is organized as follows: the ‘Method’ section is dedicated to presenting the proposed methodology and theoretical background of the article, while the ‘Material and Experimental results’ section provides details on the dataset and experiment results. The ‘Discussion’ section compares the performance results of the proposed model with previous studies. Lastly, the ‘Conclusions’ section provides the final thoughts of the article.

Pre-processing

In the pre-processing phase, we focused on enhancing the images’ quality and increasing their resolution. To achieve this goal, we utilized the very-deep super-resolution neural network, a deep learning model capable of generating high-resolution images from low-resolution images (Kim, Lee & Lee, 2016). This model leverages low-level and high-level features through a skip link to improve performance. It is specifically designed for image super-resolution tasks, which involve generating high-resolution images from low-resolution ones. The very-deep super-resolution neural network is a type of convolutional neural network (CNN) that can learn the mapping between low-resolution and high-resolution images (Kim, Lee & Lee, 2016; Ooi & Ibrahim, 2021).

The architecture of a very-deep super-resolution neural network typically consists of multiple layers of convolutional and up sampling layers. The network receives a low-resolution image as input and generates a high-resolution output image. During training, the network learns from paired examples of low-resolution and high-resolution images and strives to minimize the discrepancy between the predicted high-resolution image and the ground truth high-resolution image (Dong, Loy & Tang, 2016; Peng et al., 2021). The key feature of the very-deep super-resolution neural network is its depth. The network typically has many layers (sometimes over 100), allowing it to capture complex and high-level features in the input image. In addition, the network is often trained using residual learning, which means that the network is designed to learn the residual between the low-resolution and high-resolution images rather than the high-resolution image directly (Kim, Lee & Lee, 2016; Ren et al., 2019). This allows the network to focus on learning the high-frequency details lost in the low-resolution image. The very-deep super-resolution neural network has been shown to achieve state-of-the-art performance on various super-resolution tasks, including single-image super-resolution and video super-resolution. It has also been used in applications such as medical imaging, remote sensing, and surveillance (Kim, Lee & Lee, 2016; Shi et al., 2016; Lim et al., 2017; Ooi & Ibrahim, 2021).

In summary, the very-deep super-resolution neural network is a deep learning model capable of producing high-resolution images from low-resolution images. Its key feature is its depth, which allows it to capture complex and high-level features in the input image, and it is often trained using residual learning to focus on high-frequency details. The VSDR architecture was employed for the current study, which involves cascading convolutional layers with dimensions of 3 × 3 × 64. Additionally, the image segment size was 41 by 41. Figure 2 presents the overall framework of this network architecture.

Deep feature extraction

Multiple layers typically comprise a convolutional neural network, including convolutional, activation functions, pooling, and fully connected layers. At the beginning of a convolutional neural network is the convolutional layer, which utilizes a set of learnable filters to process the input image (Kalchbrenner, Grefenstette & Blunsom, 2014; O’Shea & Nash, 2015; Albawi, TA & Al-Zawi, 2017). Let x be the input image, and let w be a filter. The output of a convolutional layer is given by:

Here, a and b index the spatial location of the output feature map, k indexes the filter, and c indexes the input channels. The summation over m and n represents the sliding of the filter over the input image. The “∗” represents the convolution operation.

The output of the convolutional layer is then passed through an activation function, such as the ReLU function (Ramachandran, Zoph & Le, 2017):

The output of the activation function is typically fed into a pooling layer, which reduces the dimensionality of the feature map by taking the maximum (or average) value within a local neighborhood (Bayar & Stamm, 2016). The pooling operation can be represented as follows:

Where stride is the distance between adjacent pooling regions, and m and n index the local neighborhood. The output of the pooling layer is then passed to the next convolutional layer, and the process is repeated until the final layer, which is typically a fully connected layer. Finally, the output of the last pooling layer is flattened into a vector, and the fully connected layer applies a set of learnable weights to the input vector to produce the final output scores:

where W is a matrix of weights and b is a bias term. The output scores can then be passed through a softmax function to produce a probability distribution over the possible classes (Kalchbrenner, Grefenstette & Blunsom, 2014; O’Shea & Nash, 2015). Overall, convolutional neural networks are a powerful tool for image and video recognition tasks. Their ability to automatically learn feature representations from input data has led to significant advances in computer vision research.

The pre-trained architectures that include those mentioned above convolutional neural network layers offer many benefits, including faster training times, improved accuracy, transfer learning, and reduced data requirements. These benefits can make it easier to develop and deploy machine learning models for a wide range of tasks and applications (Ağdaş, Türkoğlu & Gülseçen (2021); Turkoglu, 2021; Uzen, Turkoglu & Hanbay, 2021; Imak et al., 2022). Accordingly, we have adapted pre-trained convolutional neural networks based on transfer learning for Alzheimer’s and Parkinson’s disease detection problems. In the current study, we used three robust deep architectures with different structures, DarkNet53 (Redmon & Farhadi, 2018), DenseNet201 (Huang et al., 2017), and Xception (Chollet, 2017), and extracted low- and high-level features from MRI images using the learned weights of these architectures. Detailed information about these architectures is given in Table 2.

Table 2 was utilized to obtain pre-trained deep architectures that acted as feature extractors. Deep features were extracted using fully connected layers of deep architectures given in column 2 of Table 2. Then, these obtained features are combined and given to the input of the classifier model.

Ensemble classification

In this study, we used a classification algorithm that combines the concepts of KNN and random subspace ensemble methods. KNN is a classification algorithm that predicts the class of a new instance based on the class labels of the K nearest instances in the training set. The algorithm assumes that similar instances have similar class labels (Chen, Bi & Wang, 2006; Zhang & Zhou, 2007; Zhang et al., 2017). On the other hand, random subspace ensemble methods are classification algorithms that create multiple subsets of the original feature set and train a classifier on each subset. The idea behind this approach is to reduce the variance of the classification model by introducing diversity among the classifiers (Ho, 1998; Tremblay, Sabourin & Maupin, 2004; Demir, AM & Sengur, 2020), Önder, Dogan & Polat (2023)

In KNN based random subspace ensemble classifier, KNN, and random subspace ensemble methods are combined to create a robust classification model. The algorithm first creates multiple subsets of the original feature set using random subspace ensemble methods. Then, for each subset, the KNN algorithm is applied to obtain the class labels of the nearest neighbors. Finally, the class labels predicted by each KNN model are aggregated to obtain the final prediction. This proposed ensemble classifier is represented in Fig. 3.

The random subspace ensemble classifier based on K-nearest neighbors used in the present study is a machine learning algorithm that improves classification accuracy by combining a set of KNN classifiers, each trained on a randomly chosen subset of features, with k = 1. This algorithm aims to optimise the overall performance by combining multiple KNN classifiers, each trained on a randomly selected subset of features. This method aims to create a more powerful and generalisable classifier by increasing the diversity throughout the model, with each subset emphasising different attributes. In this way, it is possible to obtain more effective and reliable classification results in complex data sets.

Material and Experimental results

In this article, we proposed a KNN-based random subspace ensemble classifier model based on deep CNN to classify Alzheimer’s and Parkinson’s diseases. We conducted the experimental works using the MATLAB 2022b platform, utilizing a workstation computer equipped with an NVIDIA Quadro P4000 card and 32GB of RAM to perform the applications. To evaluate the models utilized in the experiment, the following metrics were used: accuracy, sensitivity, specificity, precision, and F1-score (see Eqs. (5)–(9)):

(5)${\displaystyle Accuracy=\frac{\left(TP+TN\right)}{\left(TP+TN+FP+FN\right)}}$ (6)${\displaystyle Sensitivity=\frac{TP}{\left(TP+FN\right)}}$ (7)${\displaystyle Specificity=\frac{\left(TN\right)}{\left(TN+FP\right)}}$ (8)${\displaystyle Precision=\frac{\left(TP\right)}{\left(TP+FP\right)}}$ (9)${\displaystyle F1-Score=2\ast \frac{\left(Sensitivity\ast Precision\right)}{\left(Sensitivity+Precision\right)}}$

Here, true-negative (TN), true-positive (TP), false-negative (FN), and false-positive (FP) are denoted by their respective abbreviations.

Material: Dataset

Our work involved utilizing a dataset comprising MRI images of individuals diagnosed with AD, PD, and Healthy controls. The dataset (https://www.kaggle.com/datasets/farjanakabirsamanta/alzheimer-diseases-3-class) was publicly available and a valuable resource in our study. On the other hand, the dataset was split into two subsets for our study, with 80% of the data used for training and 20% used for testing. Notably, our data augmentation methods were exclusively applied to the training set.

It has been determined that the number of images between classes in the dataset used in this study is not equal. Therefore, various data augmentation techniques, including random rotate (90), vertical flip, center crop, vertical translation ([-5 5]), contrast (1.2) and brightness (0.5) were employed to address the issue of overfitting, and stabilizing the dataset. Numerical information about the raw data and the data augmented by these methods are given in Table 3.