Deep Learning-Based Multi-Feature Auxiliary Diagnosis Method for Early Detection of Ischemic Stroke

Stroke is a cerebrovascular disease resulting from damage to cerebral blood vessels due to various causes, leading to local hypoxia or ischemic necrosis of brain tissue and corresponding neurological dysfunction. It is classified into ischemic stroke and hemorrhagic stroke, with ischemic stroke accounting for approximately 75-90% of all strokes [1, 2] and presenting extremely high disability and mortality rates. According to the latest global burden of disease study in 2019, 6.55 million people succumbed to stroke worldwide, accounting for 11.6% of total deaths, making it the second leading cause of death globally, after coronary heart disease (16.2%) [3]. In China, the stroke mortality rate constitutes about 22.3% of the total mortality rate of residents, making it the leading cause of death in the country. The number of patients with cerebrovascular disease is growing at an annual rate of roughly 8.7% [3]. Considering the irreversibility of stroke development, early treatment is crucial for prognosis [4]. Currently, the clinical diagnostic method for ischemic stroke is a qualitative assessment performed by doctors based on patients' clinical characteristics and biochemical test data, primarily relying on the medical expertise and clinical experience of physicians. However, when confronted with numerous complex disease characteristics, doctors may misdiagnose and miss diagnoses due to subjective factors or heavy workload, thereby increasing their work pressure and delaying patients' treatment. Therefore, there is an urgent need for a computer-aided diagnostic method capable of rapidly diagnosing and localizing ischemic stroke lesions to assist doctors in determining treatment plans.

With the continuous advancement of artificial intelligence in the medical field, deep learning has achieved noteworthy research results in the auxiliary diagnosis of ischemic stroke. Its advantages are primarily reflected in the automatic detection of stroke [5, 6], segmentation of stroke infarction areas [7], and prediction of complications [8]. Deep learning technology can accurately extract high-quality information from a vast amount of redundant information, enabling medical staff to diagnose and devise beneficial treatment plans more rapidly in clinical work, thereby improving the speed of treatment decision-making. However, in current research on auxiliary diagnosis of ischemic stroke, diagnostic accuracy needs enhancement. The main reason is that the selection of feature types in neural networks is relatively singular or atypical features, which affect the diagnostic accuracy. For instance, the clinical decision of ischemic stroke disease based on a single image feature [9], speech feature [10], or ultrasonic inspection feature lacks the learning of diagnosis of ischemic stroke disease by combining multiple inspection features that complement each other, in order to provide patients with an accurate diagnostic method.

In light of the fact that the current method of using clinical features and biochemical test data to qualitatively diagnose ischemic stroke is not comprehensive, this study exploits the advantages of electroencephalogram (EEG) in responding rapidly and sensitively to brain metabolism, cerebral ischemia, and hypoxia [11, 12]. It is anticipated that dynamic observation of EEG changes, such as using quantitative EEG to quantitatively analyze the relevant information in EEG, and combining "clinical indicators + quantitative EEG" along with other multi-dimensional features as the input of LSTM (long short-term memory network) neural network for training will be beneficial. Simultaneously, the attention mechanism is incorporated into the neural network model to improve the diagnostic accuracy of the auxiliary diagnosis model. The area of brain infarction is closely related to the prognosis of stroke. Clinically, the diagnosis of the location and area of ischemic stroke lesions relies on CT or MRI examination to determine the specific location of the lesions. MRI can detect early stroke within tens of minutes or even minutes after stroke onset, demonstrating higher diagnostic capability than CT scans [13]. However, MRI reading is dependent on the experience and proficiency of doctors and the vast number of MRI images, resulting in a heavy workload for physicians. Coupled with individual differences among patients, accurately locating the lesion area poses certain challenges. Based on this, the present study further investigates the precise segmentation method of stroke based on MRI images in conjunction with auxiliary diagnosis. The aim is to provide more objective and accurate information to assess the severity and prognosis of ischemic stroke patients, and to assist doctors in completing rapid assessment and lesion localization of ischemic stroke.

3.1 Problem description

The research of auxiliary diagnosis for ischemic stroke can be defined as: given datasets $D=\left\{\left(d_{i 1}, d_{i 2}\right)\right\}_{i=1}^N$ and $R=\left\{r_i \in\{0,1\}\right\}$.

where, i represents the i-th patient, D represents the diagnosis data of this patient. This paper considers diagnostic indicators for ischemic stroke, which consists of clinical indicator data and quantitative electroencephalogram (QEEG) data, where d_i1 represents the clinical indicator data of the i-th patient, d_i2 represents the QEEG data of the i-th patient. R represents the diagnosis label for each patient, taking value 0 indicates non-ischemic stroke, and 1 indicates ischemic stroke. After conversion, the input features and corresponding labels are shown in formulas (2) and (3).

$X=\left\{x_1, x_2, \cdots \cdots x_n\right\}$        (2)

$Y=\left\{y_1, y_2, \cdots \cdots y_n\right\}$        (3)

In the formula above, n represents the number of patients, x_i represents the input features of the i-th patient, y_i represents the diagnosis label for the i-th patient, $y_i \in r_i$. The target of the auxiliary diagnosis model is to obtain the mapping relationship between the input features and the diagnosis label through neural network training based on the input patient data, and finally give the result of whether it is ischemic stroke.

3.2 Auxiliary diagnosis model construction

Considering the diversity, difference and correlation of the input feature information, this paper constructs an auxiliary diagnosis model for ischemic stroke based on LSTM and multi-feature fusion. At the same time, research proves that [24] the attention mechanism can improve the model's prediction accuracy by helping the model learn the correlation information between multiple features, so an attention mechanism is considered in the LSTM model. The multi-feature bidirectional LSTM network designed in this paper is shown in Figure 1, which includes two feature extraction submodules and one disease classification submodule.

1.png

Figure 1. A multifeature bidirectional LSTM network model

(1) Clinical indicator feature extraction submodule

This module realizes the feature extraction of clinical indicator data through a bidirectional LSTM model. The data consists of 15 input indicators selected in section 2.1(1) related to ischemic stroke lesions, with each indicator dimension being 1. 32 neural units are set in the bidirectional LSTM model, and a Dropout layer is added to prevent model overfitting.

(2) QEEG data feature extraction submodule

The QEEG data features mainly contain numerical indicators. The design of the submodule consists of an input layer, bidirectional LSTM layer and Dense connection layer. The 6 numerical indicators in section 2.1(2) are concatenated to form a feature vector and input into the bidirectional LSTM model. Then the bidirectional LSTM is used for feature learning, with 32 neural units calculating forward and backward features as input and outputting doubled feature vectors, with the connection mode of bidirectional LSTM being concat. Finally, Dropout with 0.5 is used to drop neurons in the network to reduce model overfitting, and a Dense connected layer outputs the feature vector.

(3) Disease classification submodule

The disease classification submodule is first responsible for connecting the feature vectors output by each feature extraction submodule and outputting the diagnosis result by the classifier. The feature vectors output by each submodule are expanded into 3D vectors that meet the input requirements of the self-attention mechanism through Lambda, and then the feature vectors are concatenated and self-attention is applied to output a feature vector by weighting the feature vectors from different modules. The dimension is compressed by Flatten, and the result is output by two Dense connected layers. Since the classification label has been converted to binary category when input, the last Dense layer outputs the classification result.

3.3 Model evaluation and experimental results

3.3.1 Diagnosis comparison experimental settings

In order to verify the effectiveness of the network in Figure 1 for auxiliary diagnosis of ischemic stroke, the following three groups of experiments were set up for result comparison.

(1) Different multifeature extraction submodel design experimental comparison

To compare the differences in performance of different designs of the proposed feature submodels, two groups of feature diagnosis models with different combinations were constructed, referred to as LSTM-LSTM and bidirectional LSTM-bidirectional LSTM according to the feature submodels.

(2) Comparison experiments of single feature models and multi-feature fusion models

To verify the effect of the proposed dual-feature fusion model (clinical indicators + QEEG data) on feature learning in the diagnosis of ischemic stroke disease compared with single feature learning, LSTM and bidirectional LSTM diagnosis models were established for each feature and used for the diagnosis of ischemic stroke as contrast experiments with the proposed model.

(3) Comparison experiment by adding attention mechanism

The attention mechanism was added to the model in this experiment to highlight the degree of attention to features from different examinations, and the experimental results were compared.

3.3.2 Explanation of experimental evaluation standards

The experiments adopted evaluation indicators that measure the reliability of disease determination in clinical diagnosis, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value and F1_Score.

where, TP represents true positive, FP represents false positive, FN represents false negative, TN represents true negative.

(1) Accuracy

Calculating the number of samples predicted correctly/total number of samples, the calculation formula is as follows:

$A c c=\frac{T P+T N}{T P+F P+F N+T N}$        (4)

(2) Sensitivity

It measures the ability to correctly determine that a patient has ischemic stroke, and calculates how many of the true ischemic stroke positive samples are truly predicted to be ischemic stroke patients. High sensitivity indicates a low miss rate, and the calculation formula is as follows:

Sensitivity $=\frac{T P}{T P+F N}$         (5)

(3) Specificity

It measures the ability to correctly identify patients without ischemic stroke. It is calculated by determining how many of all true non-ischemic stroke negative samples are accurately predicted to have non-ischemic stroke. Higher specificity indicates a lower false positive rate, as shown in the formula:

Specificity $=\frac{T N}{T N+F P}$       (6)

(4) Positive predictive value

It calculates how many of the positive samples in the predicted samples are predicted correctly, calculates the true ischemic stroke patients among those predicted to have ischemic stroke, and the calculation formula is as follows:

$P P V=\frac{T P}{T P+F P}$       (7)

(5) Negative predictive value

It calculates how many of the negative samples in the predicted samples are predicted correctly, calculates the true non-ischemic stroke patients among those predicted to have non-ischemic stroke, and the calculation formula is as follows:

$N P V=\frac{T N}{T N+F N}$       (8)

(6) F1_Score

It balances precision and recall by calculating the harmonic mean of the two, here it is calculating the harmonic mean between PPV and sensitivity, and the calculation formula is as follows:

$F 1 \_$Score $=\frac{2 \times P P V \times \text { Sensitivity }}{P P V+\text { Sensitivity }}$        (9)

3.3.3 Experimental results of auxiliary diagnosis

(1) Experimental results of different network models

First, experiments compared the diagnosis of ischemic stroke disease using different network models, including LSTM-LSTM and bidirectional LSTM-bidirectional LSTM models. The experimental results comparison is shown in Table 3.

As shown in Table 3, the bidirectional LSTM-bidirectional LSTM model overall accuracy has increased compared to the LSTM-LSTM model, indicating that using bidirectional LSTM as the hidden layer played a role in capturing reverse sequence feature information in feature extraction, The accuracy and loss changes during the training process of the two multi-feature models are shown in Figure 2 respectively.

(2) Comparison of separate feature model and multi-feature combined model experiments

To compare the effect of fusing two types of patient feature data on the diagnosis of ischemic stroke disease, separate diagnosis models for single features were established in the experiment. The comparison results of single feature diagnosis models and multi-feature diagnosis models are shown in Table 4.

Table 3. Comparison of experimental results of different network models

Table 4. Comparison of single feature diagnosis models and multi-feature diagnosis models

Table 5. Comparison of experimental results of models with added attention mechanism

2.png

Figure 2. The accuracy and loss changes during the training process of the two multi-feature models

Table 4 reveals that in this experiment, diagnostic results for two distinct types of inspection feature data using single LSTM models, single bidirectional LSTM models and the multi-feature bidirectional LSTM-bidirectional LSTM model proposed in this paper were compared. Compared to the evaluation outcomes of single feature models and single bidirectional LSTM models, the overall performance of the multi-feature bidirectional LSTM-bidirectional LSTM model reached 80%, giving play to the advantage of multi-feature information complementing each other.

(3) Experiments adding attention mechanism for comparison

In view of the role of the attention mechanism in weighting key features, this experiment added the self-attention mechanism at the multi-feature level based on the comparison of (1) models. The experimental results comparison is shown in Table 5.

As the table above shows, by adding the self-attention mechanism to both models, compared to the sensitivity and specificity of each model under self-attention, the ability to accurately diagnose ischemic stroke and non-ischemic stroke improved. In the model proposed in this paper, the ability to correctly determine ischemic stroke disease reached 89%. The positive predictive value and negative predictive value increased overall and achieved balance, indicating that the model reduced the likelihood of clinical misdiagnosis. Under the self-attention mechanism, the overall prediction accuracy of the multi-feature model remained stable or increased, demonstrating the effectiveness of the self-attention mechanism in improving model accuracy.

Currently, ischemic stroke as the main type of stroke, the incidence rate is increasing year by year globally, with high mortality, high disability and multiple complications. Timely diagnosis and determination of treatment plans are key to slowing down the progression of the disease. Currently, clinical diagnosis and localization of lesion areas still rely on doctors manually reading images, which is inefficient, laborious and subjective. Therefore, this paper proposes to build an auxiliary diagnostic model based on deep learning and an automatic lesion segmentation algorithm to help doctors quickly and accurately customize personalized treatment plans based on diagnostic results.

In building an auxiliary diagnostic model for ischemic stroke, considering that the selection of neural network feature types affects the diagnosis accuracy, EEG data is used first due to its advantages of rapidly and sensitively reacting to cerebral ischemia, combining it with clinical indicators in feature selection. A multi-feature pattern recognition method based on "clinical indicators + quantitative electroencephalogram" is proposed to screen 23 key features for neural network training. Meanwhile, attention mechanism is introduced into LSTM to design two feature extraction submodels based on bidirectional LSTM model for different types of examination data, outputting the corresponding feature vectors of each submodule and connecting to self-attention mechanism, allocating weight vectors and inputting to fully connected layer to realize auxiliary diagnosis of ischemic stroke disease. The constructed model can learn the association information between various features to further improve the diagnostic accuracy. Meanwhile, the cascaded 3D deep residual network algorithm is proposed to be applied to the segmentation of ischemic stroke lesion areas. The residual module is introduced in the segmentation architecture to avoid the problem of gradient disappearance caused by the increase of network depth and solve the problem of difficult subdividing of lesions due to unclear features and boundaries.

Although experiments prove that the overall completion of the patient's ischemic stroke auxiliary diagnosis function and automatic segmentation of ischemic stroke lesions has been realized and high diagnostic accuracy and segmentation precision have been achieved, there are still many aspects that need to be improved and perfected. For example, the training and test data in this paper came from patient data from hospitals, which need manual collection or medical equipment, there are data types or some missing. Although data cleaning and preprocessing have been done, the corrected data may not fully represent the indicators data of ischemic stroke occurrence. At the same time, due to the limited patient sample data, the data generalization ability is still not strong enough, which may affect the accuracy of disease classification diagnosis. In the future, based on the current research, further deepening and expansion is needed, and future research work mainly includes the following two aspects:

(1) In the auxiliary diagnosis model, in order to further improve the system diagnostic accuracy while training a robust neural network with stronger generalization ability, more hospital stroke patient data is needed for training and testing, and the causes of ischemic stroke need to be studied. Medical history data and expert consultation are collected to mine more indicators that can be used for the diagnosis of ischemic stroke to form neural network input features that can comprehensively reflect the symptoms of ischemic stroke.

(2) In the automatic lesion segmentation algorithm, due to the large amount of training data, especially the large number of MRI image modes and image blocks after sampling, the computational amount during network training is relatively large, resulting in relatively slow lesion segmentation speed. Therefore, it is necessary to further optimize the network structure to improve the network performance and accelerate the segmentation speed while ensuring segmentation accuracy to make the running time more efficient to meet the clinical application of hospital systems.