Early brain stroke detection using multilayer perceptron of convolutional neural network-based residual network

Abstract

Background:

Stroke, medically known as the brain attack, refers to the stoppage or stoppage of blood from flowing into a particular region of the brain, or even from the breaking of a vessel, causing injury to and death of areas of the brain. It presents a medical emergency, with the potential of severe long-term neurological impairment, disability, and even death; thus, urgent detection and treatment are needed.

Objective:

The study aims to develop a novel Multilayer Perceptron of Convolutional Neural Network-based Residual Network (MLPCNNbRN) for early brain stroke detection, focusing on improving the accuracy and reliability of detecting subtle stroke patterns in medical images.

Methods:

The MLPCNNbRN provided resented in the context of residual connections within an architecture designed for deep network training in medical images. This allowed the overall model to learn complex relations very effectively. The system was implemented in the Python framework. Its performance was compared with other methods. The key metrics used in the evaluation were accuracy, precision, recall, and F-score.

Results:

The MLPCNNbRN model demonstrated superior performance compared to existing methods, achieving higher levels of accuracy in stroke detection. Specifically, the model improved overall accuracy, precision, recall, and F-score, showcasing its robustness in identifying subtle stroke patterns.

Conclusion:

The proposed MLPCNNbRN system enhances early brain stroke detection by extracting hierarchical features and residual network learning, offering a more accurate and reliable approach than previous methods. This system has the potential to aid medical professionals in timely diagnosis and treatment, ultimately improving patient outcomes.

Keywords

Brain stroke convolutional neural network residual network stroke detection multilayer perceptron

1. Introduction

Brain cells die due to anomalies in the cerebrovascular system or cerebral circulation, which causes brain strokes.¹ A cerebral stroke is an ailment that can be fatal and is caused by inadequate blood flow to the brain. After a stroke, the brain-afflicted area stops functioning normally, underscoring the importance of early detection for enhanced therapeutic interventions.² A stroke may occur from an abrupt interruption of blood flow to a specific brain area. The resulting disability stems from the diminished blood supply, leading to the gradual loss and eventual death of brain cells in the affected region.³ Interrupted blood supply and nutrient delivery to the brain can lead to the development of symptoms. As stated by the WHO, stroke is the primary cause of disability and death globally.⁴ The clinical manifestations of stroke typically encompass weakness, paralysis, or diminished sensation in various body parts, including the possibility of hemiplegia.⁵ Stroke has emerged as a disease characterized by exceptionally high mortality rates.⁶ Hemorrhagic and ischemic strokes are treated extremely differently, but the first clinical results can often be quite similar.⁷

Although stroke carries a high mortality rate, early diagnosis offers treatability. Detecting a stroke at an early stage can mitigate severe consequences and prevent fatalities.⁸ The utilization of Machine Learning in various domains to address challenges more expeditiously than human capabilities has garnered considerable attention, propelled by the current accessibility of affordable computing power and cost-effective memory resources.⁹ Machine learning techniques have been quite important in expediting this diagnosis and triaging to strokes, leveraging feature identification and segmentation.¹⁰ The Deep Learning approach, with its ability to comprehend the intricacies of brain imaging, has a vital role in advancing the early detection of strokes.¹¹ Artificial Intelligence with a Machine Learning Subset can leverage various imaging features, including those imperceptible to the human eye, consistently achieving high levels of accuracy.¹² Efficient treatment of stroke hinges on the critical step of early detection, and Machine Learning proves to be of significant value in facilitating this process.¹³ Because deep learning can automatically calculate features inside the deep system's convolutional layers, it is commonly used as a classification approach.¹⁴

Machine learning methods have made significant progress in the prediction of many diseases, and they can be used to determine the likelihood of a stroke. Machine learning methods can be employed to assess the probability of a stroke, benefiting from substantial advancements in predicting various diseases.¹⁵ The ML model undergoes performance testing by utilizing data from one database for training and cross-validation, and data from another database for external validation.¹⁶ In the medical field, Machine Learning is extensively employed to input data for training algorithms, aiming to achieve precise and accurate outcomes.¹⁷ Several conventional techniques were utilized, including deep convolutional neural network (DCNN), you only look once (YOLO) 5, and single-shot detector (SSD).¹⁸ Computer-aided diagnosis (CAD) based brain stroke detection and classification (CAD-BSDC).¹⁹ Synthetic Minority Over Sampling Technique (SMOTE).²⁰ But there is no proper solution. So these problems are overcome to create our proposed methodology. The following can be used to summarize the key contribution of the proposed study:

First, compile a dataset of brain scans that consists of both normal and stroke-affected images.

Standardize and normalize imaging data. To improve model generalization and boost dataset diversity, use augmentation strategies.

Identify relevant features that contribute to stroke prediction. This may include age, blood pressure, cholesterol levels, and other relevant biomarkers.

Then, the proposed framework is created as a Multilayer Perceptron of Convolutional Neural Network Residual Network (MLPCNNbRN).

Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification.

Aiming to enhance early brain stroke detection through complex relationship capture in medical images

Train the MLPCNNbRN on the preprocessed dataset, optimizing parameters for effective learning.

To determine how well the trained model generalizes, test it on a different dataset.

Apply the MLPCNNbRN to new, unseen data for early brain stroke detection. Identify potential signs of stroke based on the learned features.

Eventually, F-score, Precision, Accuracy, and Recall are used to examine the performances.

The remaining section of this research is summarized as follows, section 2 elaborates on relevant literature, the System model and problem definition are detailed in Section 3, consequently, section 4 defines a novel work, and Section 5 draws the achieved results and its comparison. Finally, the paper is concluded in Section. 6.

2. Related work

Below, we present descriptions of several recent research works related to this topic.

Brain strokes have been responsible for a significant and concerning increase in fatalities and disabilities worldwide. Deep learning models, however, are unable to provide the same degree of performance for every medical application. This leads to the proposal of a framework in this study to address the problem of automatic stroke identification based on brain CT scan data. U-Net model Korra, S., et al.²¹ is configured appropriately, and data augmentation is performed. An algorithm known as Learning-based Medical Image Processing for Brain Stroke Detection (LbMIP-BSD). This algorithm optimizes brain stroke detection performance through the utilization of a U-Net-based model in supervised learning, incorporating data.

Because brain problems are on the rise and psychological stress seems to have no boundaries these days, it is believed that brain strokes will pose a serious threat to humanity in the future. Therefore, this research suggests a novel technique for predicting the beginning of a stroke by utilizing the Adaptive Neuro-Fuzzy Inference System (ANFIS) machine learning approach Lakkshmanan, A., et al.²² Following preprocessing with data cleaning techniques, SegNet is used to partition the acquired data, and CapsuleNet is used to extract pertinent features. Predictive analytics is done using the ANFIS model.

An interruption in blood flow to the brain can result in a stroke, sometimes referred to as a brain attack. This may occur from bleeding into the brain tissue or from a blockage in the blood arteries supplying the brain. Pokorny, T., et al.,²³ has introduced support vector machines (SVMs). This study set out to evaluate the effectiveness of microwave technology in the identification of brain strokes. SVMs are inherently designed for binary classification tasks, making them suitable for distinguishing between stroke and non-stroke cases, which is often the goal in early detection.

Stroke, a prevalent neurological disorder, arises from abnormal blood flow in the brain. In addition, stroke has been leading the cause of death worldwide in recent years. In this chapter, Dandıl, E. et al.²⁴ offer a fully automated approach for segmenting brain stroke lesions using MR images that is based on a mask region-based convolutional neural network (Mask R-CNN). The results of the study indicate that the suggested Mask R-CNN method can be a useful tool for accurately separating stroke lesions with unique borders.

This article offers an automated epilepsy seizure detection system (AESD) utilizing Deep Learning (DL) algorithms Kumar, K.D., et al.²⁵ to diagnose and detect epilepsy seizures based on EEG signals, which are wavelet signals widely used for identifying abnormal brain events. The dataset encompasses attributes such as patient family history, age, and details of long-term medication usage, addressing various challenges associated with epilepsy diagnosis. Performance evaluation is conducted based on relevant parameters.

Although they are mostly curable and avoidable, cerebrovascular disorders like stroke are among the leading causes of death and disability worldwide. Recognizing strokes early and promptly intervening are pivotal in lessening the disease burden and enhancing clinical outcomes. In recent times, the spotlight has turned to machine learning methodologies for stroke detection due to their potential. Saleem, MA, et al.,²⁶ to achieve this objective, an early stroke detection system leveraging CT brain images, alongside a genetic algorithm and a Bidirectional long short-term memory (BiLSTM) model, to identify strokes at their nascent stages. For image classification, we employed a genetic algorithm-driven neural network approach to identify the most pertinent features for classification, which were subsequently fed into the BiLSTM model.

To improve communication for people with brain stroke or paralysis, this research combines Independent Component Analysis (ICA) with several deep learning algorithms in the Brain-Computer Interface (BCI) system, Kanimozhi, M., et al.²⁷ Using EEG data from BCI competition III, it explores classification methods that leverage ICA-extracted components and various deep learning architectures. The innovation is in figuring out what the essential elements are and incorporating them into the network and the comparison study compares results with and without dimensionality reduction by ICA. LSTM, GRU, and CNN are a few instances of deep learning models designed to outperform current techniques in terms of EEG signal processing and classification accuracy. Research gap of the existing works are demonstrated in Table 1.

3. System model and problem statement

A stroke can result from an interruption in blood flow to a specific area of the brain, which can have permanent consequences, including damage to brain tissue. There are two primary stroke kinds: hemorrhagic stroke, in which brain bleeding and increased intracranial pressure result from a blood vessel burst, and ischemic stroke, which results from a blockage in a blood vessel, depriving the affected brain tissue of oxygen and nutrients. Both types of stroke lead to neurological symptoms such as sudden weakness, numbness, difficulty speaking, and vision problems, which are demonstrated in Figure 1. If not treated promptly, strokes can lead to brain tissue death and permanent disability. Early recognition and medical intervention are crucial for improving outcomes and minimizing long-term effects.

Figure 1.

System model & problem statement.

The brain serves as the central organ in the human body. A medical disorder known as the brain's blood arteries burst, which results in a stroke, it causes harm to the brain. Interruption of blood supply can result in symptoms, and stroke is considered a medical emergency due to its risk of potential mortality, complications, and long-term brain damage. The WHO has identified stroke as a primary global cause of mortality and disability. Therefore, minimizing severity and optimizing outcomes require early diagnosis of stroke warning symptoms. These issues have motivated me to develop this proposed work.

4. Proposed methodology

The present research article has aimed to design a novel MLPCNNbRN. Early brain stroke detection using MLPCNNbRN involves Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification. Figure 2 shows the architecture flow of the proposed methodology.

Figure 2.

Proposed methodology for MLPCNNbRN.

First, compile a dataset of brain scans that consists of both normal and stroke-affected images. Standardize and normalize imaging data. To improve model generalization and boost dataset diversity, use augmentation strategies. Identify relevant features that contribute to stroke prediction. This may include age, blood pressure, cholesterol levels, and other relevant biomarkers. Then, the proposed framework is created as a MLPCNNbRN. Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification. Aiming to enhance early brain stroke detection through complex relationship capture in medical images. Train the MLPCNNbRN on the preprocessed dataset, optimizing parameters for effective learning. To determine how well the trained model generalizes, test it on a different dataset. Apply the MLPCNNbRN to new, unseen data for early brain stroke detection. Identify potential signs of stroke based on the learned features. Eventually, F-score, Precision, Accuracy, and Recall are used to examine the performances.

4.1. Data set

We have used Brain stroke images in this methodology. The brain stroke dataset features two main categories: “stroke_cropped” and “stroke_noncropped,” each with specific testing, training, and validation subsets. With images cropped to focus on key areas and original non-cropped images provided, the dataset, at 73.4 MB, is invaluable for stroke-related image analysis. Researchers and practitioners can explore diverse experiments, catering to different processing needs, and ensuring effective model training and evaluation. These datasets are crucial for improving stroke diagnosis and treatment through advanced technology.

\begin{aligned} D = {(x_{i}, y_{i})}_{i = 1}^{N} \end{aligned}

(1)

Where D is denoted as data gathering,

x_{i}

represents the brain image data,

y_{i}

and represents the corresponding label (Stroke-affected or normal).

4.2. Pre-processing

In early brain stroke detection preprocessing using deep learning, standardizing and normalizing imaging data involves ensuring consistent pixel values and scaling to a standard range. Augmentation techniques are applied to increase dataset diversity, such as rotating, flipping, or zooming images, enhancing model generalization. To successfully learn from and extract relevant information from brain images, preprocessing techniques are essential for deep learning algorithms. This improves the accuracy of stroke detection and facilitates prompt medical interventions.

P_{x} = \frac{x_{i} - x_{m i n}}{x m i n_{m a x}}

(2)

Where

x_{i}

is the Original value,

P_{x}

what is the normalized value,

x_{m a x}

and

x_{m i n}

what are the maximum and the minimum values.

4.3. Feature extraction

Feature extraction in early brain stroke detection plays a crucial role in identifying relevant information from brain imaging data that contributes to predicting strokes. These characteristics usually include a variety of variables, including age, cholesterol, blood pressure, and different biomarkers that are taken out of patient files.

F = P (σ (W_{3} . P (σ (W_{2} . P (σ (W_{1} . I + b_{1})) + b_{2})) + b_{3}))

(3)

Where F represents the extracted features, I denotes the input medical image data, Weight matrices are

W_{1}, W_{2}, W_{3}

, bias vectors are

b_{1}, b_{2}, b_{3}

, Pooling layers P to extract hierarchical features from the image

σ

is denoted as activation function.

4.4. Proposed MLPCNNbRN

Early brain stroke detection using a MLPCNNbRN involves a comprehensive approach to analyzing medical images for subtle signs of stroke. Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification. Aiming to enhance early brain stroke detection through complex relationship capture in medical images. The Multilayer Perceptron (MLP) component extracts hierarchical features, capturing intricate details and abstract patterns from the images. The Convolutional Neural Network (CNN) segment then identifies spatial features relevant to stroke, while the Residual Network (ResNet) principles enhance training by mitigating gradient issues in deep networks. By jointly learning complex relationships within the data, MLPCNNbRN excels at accurately identifying potential signs of stroke, enabling healthcare professionals to intervene early and improve patient outcomes.

4.4.1. Multilayer perceptron (MLP)

Utilizing early brain stroke detection entails leveraging its deep learning capabilities to extract meaningful features from medical images, distinguishing between healthy and stroke-affected brain patterns. Through training on labeled data, the MLP learns to identify subtle variations indicative of strokes, aiding healthcare professionals in timely intervention and treatment planning. The MLP consists of multiple hidden layers, each with a set of neurons. The following formula is used to determine each neuron's output:

Z_{i} = f (W_{i} . X + b_{i})

(4)

Where

Z_{i}

denotes the output of the neuron i, the activation function f,

W_{i}

is the weight matrix for the neuron i,

b_{i}

shows the bias vector for the neuron i, and X represents the input medical image data.

4.4.2. CNN-based resNet

The proposed CNN is a bio-inspired neural network that combines convolutional layers, pooling, and a fully connected network akin to a Multi-layer Perceptron. As a counterpart to the traditional feed-forward network in image processing, it embodies three distinct layers: input, hidden, and output, with units depicting neurons in each layer. The proposed CNN architecture will consist of a 1D convolutional layer, a 1D pooling layer, and a fully connected layer. The convolution layer of the proposed CNN extracts features of the input 1D sequence data through sliding filters and output feature maps. Considering hyperparameter optimization, it has selected the number of filters and the length of the filters to be used. Nonlinear activation functions will be used in both input and hidden layers, but ResNet is used except for output. This is because the ResNet function solves problems associated with vanishing gradients and errors. Additionally, the learning process is fast. Mathematically, ResNet is represented as Equation (5).

f (y) = max (0, y)

(5)

Where y is denoted the input, A method of maximum 1D pooling was proposed to reduce the dimensionality of features.

ResNet has been chosen for the proposed work since it doesn't suffer from gradient difficulties that plagued several deep learning models. A CNN network suffers from a fading gradient problem during training. An earlier layer's gradient rule declined to zero or nil as learning progressed. The problem can be solved through the ResNet technique. The second inner layer of ResNet will now arise by adding up all of its direct inputs and further with a residual layer that gets an insight into the connection between those components.

Early brain stroke detection using a CNN-based ResNet harnesses deep learning's power for intricate feature extraction from medical images, vital for spotting subtle stroke indications early. ResNet's residual connections aid in training deeper layers effectively, improving model performance by capturing complex spatial relationships. By training on a diverse dataset, the CNN-based ResNet learns to accurately distinguish between normal and stroke-affected brain images, facilitating prompt intervention and tailored treatments for at-risk patients. The output of a convolutional layer can be represented as:

\begin{aligned} H = f (W_{c o n} * Z_{i} + b_{c o n}) \end{aligned}

(6)

Where H indicates the result of the CNN layer,

W_{c o n}

is the convolutional filter weights,

Z_{i}

represents the output from the MLP layer, and

b_{c o n}

denotes the bias term.

The ResNet architecture includes residual connections to facilitate training deeper layers effectively. Let's denote the result of the CNN layer as H. The result of a residual block in ResNet can be represented as:

H_{r e s} = f (W_{r e s} * H + b_{r e s}) + H

(7)

Where

H_{r e s}

is the output of the residual block,

W_{r e s}, b_{r e s}

are the weights and bias the weights and bias of the residual block, H is the input to the block, and f is the activation function.

4.5. Training & testing

After training the MLPCNNbRN model, it can be used to identify strokes early on in the process of analyzing newly acquired brain pictures. The model uses the features it learned during training to classify these new images as either normal or stroke-affected. The testing set is a different collection of brain images that we utilize to analyze the model's functionality. We feed these unseen images through the trained model to get predictions. By setting a threshold on the model's output probabilities, we can determine if an image is classified as normal or stroke-affected. This process helps generate detection results and assess the model's accuracy in identifying strokes, which is crucial for determining its effectiveness in real-world scenarios.

L (θ) = \frac{1}{N} \sum_{i = 1}^{N} L (y_{i}, x_{i})

(8)

Where L is the loss function,

θ

represents the network parameter, N shows the number of training samples,

y_{i}

represents the true labels,

x_{i}

and indicates the predicted label.

Y_{t e s t} = σ (W_{x} . F_{f c} + b_{x})

(9)

Where

Y_{t e s t}

the predicted probability of stroke occurrence during testing,

σ

sigmoid activation function for binary classification,

W_{x}

is Weight matrix,

F_{f c}

Output of the fully connected layer, combining extracted features,

b_{x}

Bias vector.

4.6. Detection

Early brain stroke detection using an MLPCNNbRN involves training a sophisticated computer model to recognize subtle signs of strokes in brain images. This model combines the power of a Multilayer Perceptron (MLP) for classification with a CNNbResNet for feature extraction, allowing it to accurately classify brain scans as normal or indicative of a stroke. This strategy seeks to enhance early stroke detection, which is crucial for prompt medical intervention additionally improved patient outcomes. It does this by utilizing deep learning techniques.

P (Y = 1 / X_{t e s t}) = σ (W . H_{r e s} + b_{i})

(10)

Where,

P (Y = 1 / X)

represents the predicted probability of stroke presence, W is the weight for the prediction layer,

b_{i}

is the bias term,

σ

and is the sigmoid activation function.

If $P (Y = 1 / X_{t e s t}) > θ$ , Classify the image as stroke-affected (1)

If $P (Y = 1 / X_{t e s t}) \leq θ$ so, Classify the image as normal (0).

The binary classification output is based on applying the threshold. If the predicted probability

P (Y = 1 / X)

is greater than the threshold

θ

, the image is classified as stroke-affected (1), otherwise, it is classified as normal (0).

Advanced machine learning techniques are utilized with the MLPCNNbRN to enhance the speed along with accuracy in stroke diagnosis. In this hybrid approach, the capability of CNNs for feature extraction is fused with the robust classification power of the MLP and residual learning framework that helps to train much deeper networks by actually overcoming the vanishing gradient problem. The MLPCNNbRN might be very effective at picking up very subtle patterns from processing medical imaging data, such as MRI or CT scans, that could indicate a stroke, allowing for timely medical intervention. This integration further enhances the detection process with an ultimate focus on reducing time to treatment and improving patient outcomes within the emergency healthcare environment.

To begin the process of early brain stroke detection, a dataset comprising brain images, including samples from both stroke-affected and normal brains, is gathered. These images undergo preprocessing steps such as standardization and normalization to ensure consistency and remove any biases in pixel values. Feature extraction techniques are then applied to the preprocessed images to capture relevant information that contributes to stroke prediction. The proposed model is an MLP of a CNN-based Residual Network, which combines the hierarchical feature extraction capabilities of CNNs with the depth and skip connections of ResNets. The extraction of intricate hierarchical information from medical images is made possible by this design, which helps in the early identification of brain strokes. To gauge how well the model performs in correctly identifying stroke patterns, a portion of the dataset is used for training, and another set is used for evaluation. The purpose of this process is to use modern deep-learning methods to create a reliable and accurate system for the early identification of brain strokes. The suggested MLPCNNbRN flow chart is shown in Figure 3.

Figure 3.

Flow chart of proposed MLPCNNbRN.

5. Results and discussion

The Python framework has been utilized in this study to implement the suggested MLPCNNbRN approach. The proposed model incorporates the features of a Multilayer Perceptron of Convolutional Neural Network based Residual Network. Early brain stroke detection using MLPCNNbRN involves Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification. We have used Brain stroke images in this methodology. In this research, we used the Windows 10 version. The experimental specifications are mentioned in Table 2.

5.1. Performance estimation

The proposed model is intended to run within the Python framework. The precision, accuracy, F-score, and recall of the model were compared to those of other models to validate its efficiency score. Moreover the existing techniques like Radial Basis Function Networks (RBF), Long short-term memory (LSTM),³ SVM,²⁸ and convolutional deep belief networks (CDBN),²⁹ Logistic Regression (LR),⁴ Convolutional Neural Network (CNN).³⁰

5.1.1. Comparison of accuracy (a)

The model's total forecast accuracy is measured by accuracy. The ratio between the entire count of occurrences inside the dataset and the accurately predicted instances (true positives and true negatives) is determined.³¹ When it comes to early brain stroke diagnosis, accuracy is the model's ability to correctly forecast whether a given instance is normal or stroke-affected. It is a crucial metric that assesses how accurate the model is overall in predicting both classes. Mathematically, accuracy is given by:

A = \frac{T P + T N}{T P + T N + F P + F N}

(11)

Where, the true positive rate (TPR) is indicated as TP, the true negative rate is indicated as TN, the false positive rate is FP, and the false negative rate is represented as FN.

Accuracy calculations are used to validate the effectiveness of the suggested model MLPCNNbRN approach. Figure 4. Comparison of the existing method in terms of Accuracy gained 90.00%, RBF attained 82.09% accuracy, LSTM pertained 88.03% accuracy measure, SVM method attained 87.12% accuracy and the technique CDBN achieved 89.02% accuracy rate.

Figure 4.

Comparison of existing methods in terms of Accuracy.

5.1.2. Comparison of precision (p)

Out of all the cases that the model predicts to be stroke cases, the precision metric quantifies the proportion of actual stroke cases. It aids in comprehending how the model prevents false positives.³² Out of all the instances that were predicted to be positive, this is the percentage of actual stroke cases that the model correctly recognized. It can be determined as follows:

\begin{aligned} P = \frac{T P}{T P + F P} \end{aligned}

(12)

When using the developed MLPCNNbRN approach, precision refers to the estimation of evaluating the accurate detection. The effectiveness of the created MLPCNNbRN is demonstrated by comparison with other methods, including RBF, LSTM, SVM, and CDBN. RBF achieved 77.86% of precision measure, LSTM attained 86.88% as the precision, SVM achieved 85.67% of precision measure and the CDBN method earned 89.16% precision measure. As a result, a precision rate of 92.00% was achieved by the suggested method. Figure 5. Comparison of existing method in terms of Precision.

Figure 5.

Comparison of existing method in terms of Precision.

5.1.3. Comparison of recall (r)

Recall measures the model's capacity to accurately separate positive examples from all real positive instances. It is often referred to as sensitivity or TPR.³³ Recall quantifies the model's ability to correctly identify all positive instances. Specifically, it represents the proportion of actual stroke-affected images that were correctly classified by the model. A high recall is crucial in medical diagnostics, as it ensures that most cases of stroke are detected, thus enabling timely intervention and treatment. The ratio of all real positive occurrences to all true positive forecasts (false negatives and true positives) is computed. Mathematically, recall is given by:

R = \frac{T P}{T P + F N}

(13)

Where FN represents the false negative rate and TP denotes the TPR.

Figure 6. Recall the comparison of the current approach. The MLPCNNbRN technique that was developed achieved a recall measure of 90.00%. As a result, the RBF of the current method was 74.52%, the LSTM approach's recall measure was 81.31%, and the SVM method's recall measure was 86.89%. In the end, the CDBN method produced a recall measure of 86.71%.

5.1.4. Comparision F-score (F)

The F1-score is the harmonic mean of precision and recall. It provides a balance between the two metrics, offering a single score that represents the model's performance in both identifying true positives and minimizing false positives. This metric is particularly useful when you want to maintain a balance between precision and recall, especially in cases where one may be prioritized over the other.¹⁰ The F-score is calculated using the formula below:

F = \frac{2 \times P \times R}{P + R}

(14)

The created MLPCNNbRN model's F-score value is verified using established methods such as RBF, LSTM, SVM, and CDBN. The comparison of the current approach in terms of F-score (F) is shown in Figure 7. The estimated F-score of the suggested approach is 91.00%, which is more efficient than the methods that are currently in use. Accordingly, the existing method RBF attained 74.97%, the F-score of the LSTM approach was 83.28%, SVM method attained an 86.34% F-score. Finally, the CDBN technique achieved a 90.32% F-score.

Figure 6.

Comparison of existing method in terms of F-Score.

Figure 7.

Comparison of existing method in terms of Recall.

5.2 Discussion

The overall performance of early brain stroke detection using MLPCNNbRN is highly promising, offering accurate stroke classification with strong precision, recall, and F1 scores. The model benefits from the deep learning architecture, effectively capturing complex image patterns. However, its high computational requirements and potential interpretability issues limit its practical use in some real-world clinical settings.

Table 3: Comparing the estimation of overall performance as measured by F-score, recall, accuracy, and precision. Accuracy calculations are used to validate the effectiveness of the suggested model MLPCNNbRN approach. As a result, the accuracy of the suggested model increased to 90.00%. As a result, a precision rate of 92.00% was accomplished using the recommended technique. The MLPCNNbRN technique that was developed achieved a recall measure of 90.00%. Compared to the current methods, the suggested method's F-score estimation is more efficient, having reached 91.00%.

Table 1.
Research gap of previous studies.

SI.NO Author name Techniques Merits Demerits

1 Korra, S., et al.²¹ U-Net model Supervised learning with U-Net and data augmentation enhances brain stroke detection performance. The complexity inherent in the traditional U-Net architecture may impact both training efficiency and accuracy.

2 Lakkshmanan, A., et al.²² ANFIS Data undergoes preprocessing, segmentation, and feature extraction with SegNet and CapsuleNet, followed by predictive analytics using ANFIS. Complexity in model design and tuning.

3 Pokorny, T., et al.²³ Support Vector Machines(SVMs). SVMs excel at binary classification, ideal for distinguishing stroke from non-stroke cases in early detection efforts. It is not suitable for large data sets.

4 Dandıl, E. et al.²⁴ Mask R-CNN Accurate stroke region segmentation and localization. Combines segmentation and object detection into a single framework. High computational complexity. Requires large annotated datasets.

5 Kumar, K.D., et al.²⁵ Automated epilepsy seizure detection system (AESD) Early seizure identification allows for timely medical attention. Limited interpretability of complex algorithms, hindering clinical decision-making.

6 Saleem, MA, et al.²⁶ Bidirectional long short-term memory (BiLSTM) Bidirectional context understanding in BiLSTMs enhances the model's ability to detect complex patterns from past and future information simultaneously. BiLSTM is a much slower model and requires more time for training.

7 Kanimozhi, M et al.²⁷ Brain-Computer Interface (BCI) It allows individuals to manipulate an external device using signals from their brain. BCI records brain responses that are inconsistent across time and situations.

SI.NO	Author name	Techniques	Merits	Demerits
1	Korra, S., et al.²¹	U-Net model	Supervised learning with U-Net and data augmentation enhances brain stroke detection performance.	The complexity inherent in the traditional U-Net architecture may impact both training efficiency and accuracy.
2	Lakkshmanan, A., et al.²²	ANFIS	Data undergoes preprocessing, segmentation, and feature extraction with SegNet and CapsuleNet, followed by predictive analytics using ANFIS.	Complexity in model design and tuning.
3	Pokorny, T., et al.²³	Support Vector Machines(SVMs).	SVMs excel at binary classification, ideal for distinguishing stroke from non-stroke cases in early detection efforts.	It is not suitable for large data sets.
4	Dandıl, E. et al.²⁴	Mask R-CNN	Accurate stroke region segmentation and localization. Combines segmentation and object detection into a single framework.	High computational complexity. Requires large annotated datasets.
5	Kumar, K.D., et al.²⁵	Automated epilepsy seizure detection system (AESD)	Early seizure identification allows for timely medical attention.	Limited interpretability of complex algorithms, hindering clinical decision-making.
6	Saleem, MA, et al.²⁶	Bidirectional long short-term memory (BiLSTM)	Bidirectional context understanding in BiLSTMs enhances the model's ability to detect complex patterns from past and future information simultaneously.	BiLSTM is a much slower model and requires more time for training.
7	Kanimozhi, M et al.²⁷	Brain-Computer Interface (BCI)	It allows individuals to manipulate an external device using signals from their brain.	BCI records brain responses that are inconsistent across time and situations.

Table 2.

Experimental setup.

Parameter description
Processor	Intel(R) Core(TM) i5-3570 CPU @ 3.40 GHz 3.40 GHz
Product ID	00331-10000-00001-AA421
Installed RAM	4.00 GB (3.88 GB usable)
Device ID	184401E8-2212-4AFD-A513-1421313E3560
Software used	Jupyter Notebook
System type	64-bit operating system, x64-based processor

Table 3.

Comparison of overall performance estimation.

Methods	Performance Estimation
Methods	Accuracy (%)	Precision (%)	Recall (%)	F-score (%)
RBF	82.09	77.86	74.52	74.97
LSTM	88.03	86.88	81.31	83.28
SVM	87.12	85.67	86.89	86.34
CDBN	89.02	89.16	86.71	90.32
LR	71.00	69.00	73.00	71.00
CNN	81.00	76.00	82.00	73.00
Proposed MLPCNNbRN	90.00	92.00	90.00	91.00

Early brain stroke detection using a MLPCNNbRN faces several limitations that can impact its effectiveness and implementation in clinical settings. One major limitation is data dependency, as the model's performance is heavily influenced by the quality and quantity of training data; inadequate or imbalanced datasets may lead to poor generalization and increased false positives or negatives. Additionally, the complexity of the model raises concerns about overfitting, particularly when the training dataset lacks sufficient diversity, resulting in a system that performs well on training data but fails on unseen cases. Interpretability is another challenge, as deep learning models often operate as black boxes, making it difficult for clinicians to understand their decision-making processes, which can hinder trust in the technology. Moreover, MLPCNNbRN requires substantial computational resources for training and inference, potentially limiting access in resource-constrained healthcare environments. Variability in imaging conditions, such as differences in equipment and patient positioning, can further affect performance consistency. Finally, regulatory and ethical concerns surrounding patient privacy, data security, and compliance with medical standards may complicate the integration of AI in medical diagnostics, delaying the deployment of such advanced stroke detection technologies.

6. Conclusion

Numerous neurological impairments, including paralysis or weakness on a single body side (hemiparesis or hemiplegia), difficulty comprehending or speaking (aphasia), vision impairment, issues with coordination, and sensory difficulties, can result from a brain stroke. To overcome this issue a novel MLPCNNbRN. Early brain stroke detection using MLPCNNbRN involves Extracting hierarchical features from medical images, leveraging residual connections for deep network training, and jointly learning complex relationships, enhancing accuracy in subtle stroke pattern identification. Future research will focus on integrating multi-modal data sources, including imaging data, clinical data, genetic data, and real-time physiological monitoring data. In this scenario, the developed model has achieved impressive outcomes, boasting a remarkable 90.00% accuracy, coupled with a high precision of 92.00%, 90.00% recall rate, and 91.00% is an F-score. As a result, the performance of the suggested model was contrasted with that of current techniques, verifying its efficacy in achieving improved outcomes and displaying superior results. In the future, the developed work will further improve the validation and training and increase the model's generalizability across various populations, using larger and more varied datasets. Also, healthcare professionals are incorporating with this model into the current medical imaging systems to streamline processes.

Footnotes

Ethical approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Consent to publish

All the authors involved in this manuscript give full consent for publication of this submitted article.

Consent to participate

All the authors involved have agreed to participate in this submitted article.

Authors contributions

All authors have equal contributions in this work.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

Data sharing does not apply to this article.

References

Turovsky

Baryshev

Plotnikov

. Selenium nanoparticles in protecting the brain from stroke: possible signaling and metabolic mechanisms. Nanomaterials 2024; 14: 160.

Ozaltin

Coskun

Yeniay

, et al. A deep learning approach for detecting stroke from brain CT images using OzNet. Bioengineering 2022; 9: 783.

Samman Ashraf

Butt

Sharif

. Brain stroke prediction using learning machines & deep learning, 2023.

Rahman

Hasan

Sarkar

. Prediction of brain stroke using machine learning algorithms and deep neural network techniques. Eur J Electric Eng Comput Sci 2023; 7: 23–30.

Cao

Song

, et al. Deep learning derived automated ASPECTS on non-contrast CT scans of acute ischemic stroke patients (Vol. 43, No. 10, pp. 3023–3036). Hoboken, USA: John Wiley & Sons, Inc, 2022.

Wang

, et al. Classification and location of cerebral hemorrhage points based on SEM and SSA-GA-BP neural network. IEEE Transactions on Instrumentation and Measurement, 2024.

Wei

Liu

, et al. Deep learning for hemorrhagic lesion detection and segmentation on brain CT images. IEEE J Biomed Health Inform 2021; 25: 1646–1659.

Tursynova

Omarov

Tukenova

, et al. Deep learning-enabled brain stroke classification on computed tomography images. Comput Mater Continua 2023; 75: 1431–1446.

Khan

Kader

Islam

, et al. Machine learning and deep learning approaches for brain disease diagnosis: principles and recent advances. IEEE Access 2021; 9: 37622–37655.

10.

Sherif

Ahmed

. A machine learning approach for stroke differential diagnosis by blood biomarkers. J Adv Inf Technol 2024; 15: 1–9.

11.

Omarov

Tursynova

Uzak

. Deep learning enhanced internet of medical things to analyze brain computed tomography images of stroke patients. Int J Adv Comput Sci Appl 2023; 14.

12.

Lee

Ham

, et al. Machine learning approach to identify stroke within 4.5 hours. Stroke 2020; 51: 860–866.

13.

Srinivas

Mosiganti

. A brain stroke detection model using soft voting-based ensemble machine learning classifier. Measur Sensors 2023; 29: 100871.

14.

Gautam

Raman

. Towards an effective classification of brain hemorrhagic and ischemic stroke using CNN. Biomed Signal Process Control 2021; 63: 102178.

15.

Mostafa

Elzanfaly

Yakoub

. A machine learning ensemble classifier for prediction of Brain Strokes. Int J Adv Comput Sci Appl 2022; 13.

16.

Alruily

El-Ghany

Mostafa

, et al. A-Tuning ensemble machine learning technique for cerebral stroke prediction. Appl Sci. 2023; 13: 5047.

17.

Mahmud

Mamun

Abdelgawad

. A deep analysis of brain tumor detection from MR images using deep learning networks. Algorithms 2023; 16: 176.

18.

Sailaja

Pattani

. Effective brain stroke prediction with deep learning model by incorporating YOLO_5 and SSD. Int J Online Biomed Eng 2023; 19: 35754–35764.

19.

Eshmawi

Khayyat

Algarni

, et al. An ensemble of deep learning enabled brain stroke classification model in magnetic resonance images. J Healthc Eng. 2022; 2022: 1–11.

20.

Kadhim

Radhi

. Machine learning prediction of brain stroke at an early stage. Iraqi J Sci. 2023; 12: 6596–6610.

21.

Korra

Soora

Jahan

, et al. Brain CT image processing using U-net model with data augmentation for detection of ischemic and haemorrhage strokes. Int J Intell Syst Appl Eng 2024; 12: 72–82.

22.

Lakkshmanan

Priyanka

Anand

, et al. A novel machine learning based stroke prediction system using magnetic resonance imaging and adaptive new fuzzy inference system. Int J Intell Syst Appl Eng 2024; 12: 576–585.

23.

Pokorny

Vrba

Fiser

, et al. On the role of training data for SVM-based microwave brain stroke detection and classification. Sensors 2023; 23: 2031.

24.

Dandıl

Yıldırım

. Fully automated segmentation of brain stroke lesions using mask region-based convolutional neural network. In: Diagnosis of neurological disorders based on deep learning techniques. CRC Press, 2023, pp.113–130.

25.

Kumar

Dash

Ganiya

. Detection of epileptic seizures using hybrid deep learning approaches. Int J Intell Syst Appl Eng 2024; 12: 174–180.

26.

Saleem

Javeed

Akarathanawat

, et al. Innovations in stroke identification: a machine learning-based diagnostic model using neuroimages. IEEE Access 2024; 64.

27.

Kanimozhi

Roselin

. Enhancing motor imagery classification with deep learning: a focus on dimensionality reduction and temporal features. Int J Intell Syst Appl Eng 2024; 12: 121–128.

28.

Shinde

Kurhekar

Gulhane

, et al. Design of a novel enhanced machine learning model for early prediction of cerebral stroke. Int J Comput Digit Syst 2024; 16: 1–22.

29.

Zebari

Mohammed

Zebari

, et al. A deep learning fusion model for accurate classification of brain tumors in Magnetic Resonance images. CAAI Trans Intell Technol 2024; 9: 790–804.

30.

Karakis

Gurkahraman

Mitsis

, et al. Deep learning prediction of motor performance in stroke individuals using neuroimaging data. J Biomed Inform. 2023; 141: 104357.

31.

Paralic

Zelenak

Kamencay

, et al. Automatic approach for brain aneurysm detection using convolutional neural networks. Appl Sci. 2023; 13: 13313.

32.

Uddin

Kumar Sarker

Hoque

, et al. An explainable machine learning framework for early detection of stroke. Int Multiling J Sci Technol 2023; 8.

33.

Ushasree

Krishna

Rao

, et al. SPE: ensemble hybrid machine learning model for efficient diagnosis of brain stroke towards clinical decision support system (CDSS). Int J Intell Syst Appl Eng 2023; 11: 339–347.