Learning Methods of Convolutional Neural Network Combined With Image Feature Extraction in Brain Tumor Detection

Computer-aided detection technology is less applied in brain tumor detection terminals, and it is difficult to eliminate the influence of various interference factors on the diagnosis results. In order to promote the application of computer-aided detection technology in brain tumor detection, this study based on convolutional neural network, combined with MRI detection technology to construct a model adapted to brain tumor feature detection. The main function of this research model is to segment and recognize MRI brain tumors and use convolutional layer to perform convolution operation to improve recognition efficiency and rate and combine artificially selected features with machine learning features. In addition, this article uses feature fusion to further improve the diagnostic results. Finally, this article designs experiments to perform performance analysis. The research shows that the model algorithm designed in this article has certain practical effects and can provide theoretical reference for subsequent related research.


I. INTRODUCTION
In recent years, the level of diagnosis and treatment of brain tumors is constantly improving, and imaging techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) are also widely used in brain tumor detection. These techniques are very effective in examining patients with brain tumors, and the detection rate is also relatively high. MRI has become an important part of modern imaging medicine. According to relevant clinical research, MRI is better than CT in the diagnosis of intracranial brain tumors, which can reach 98% correct rate [1]. MRI imaging technology has many advantages: on the one hand, people are more concerned about radiation, but this imaging technology will not cause harm to the human body. On the other hand, it can be imaged with multiple parameters, and this imaging approach can provide a wealth of useful information for diagnostic purposes, and it is also more convenient and effective for studying human metabolism and function. In addition, MRI imaging technology provides a wealth of anatomical information on human soft tissue [2]. In medicine, MRI imaging The associate editor coordinating the review of this manuscript and approving it for publication was Zhihan Lv . technology is generally used for brain tumor segmentation and detection, and experts analyze the MR images to determine the existence and development of brain tumors [3].
QBIC (QueryByImageContent) [4] is a commercial content-based video and image retrieval system developed by IBM. It mainly includes the function of retrieving dynamic video and static images. Moreover, it also supports text-based retrieval, supports sketch retrieval, supports mixed retrieval based on text and content, and retrieval based on contentbased retrieval methods. The QBIC system uses the K-L transform to reduce the dimensionality design and highdimensional feature indexing techniques for multidimensional features and adopts the pattern recognition idea for the database [5]. The QBIC image retrieval system is used earlier in actual retrieval applications, and its robustness is good. The function of its humanized retrieval system satisfies the basic retrieval requirements of searchers, and its design ideas and framework structure lay an important foundation for the development and verification of content-based image retrieval technology in the future. The development of QBIC is inseparable from the continuous improvement and verification of a large number of researchers, so that it has been widely used in life and practical teaching. For example, VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ the University of California introduced the QBIC system to allow parents and students to check the teacher's information in a timely manner and was welcomed by everyone [6]. The VIRImageEnginet, developed by the Massachusetts Institute of Technology and the University of Michigan, supports an extensible image retrieval system. The system is mainly composed of texture, shape and color features. We can make these features randomly combined, and the searcher can adjust the weights between different features according to the results of the first search. After the adjustment, the search is performed again, and the weights are slightly adjusted, then finally the search results expected by the searcher can be obtained [7]. ASSERT, 9, designed by Purdue University, is a retrieval system based on lung CT images. It uses feature extraction algorithm integration, can automatically adapt to different types of lung disease images for feature extraction, and has been effectively verified in clinical. The innovation is that the shape feature extraction algorithm of this system uses the automatic segmentation extraction contour algorithm [8].
Brain tumors show relatively slow clinical symptoms, and the duration of onset varies. Brain tumors include different types of tumor problems: solid or active brain tumors, edema, and necrosis. Early symptoms are common in the early morning headache, which usually occurs at 4 or 5 in the morning; visual impairment is also one of the early symptoms of brain cancer, because the increase in intracranial pressure causes poor venous return of the eyeballs, which forms edema and weakens vision; In severe patients, the retina will have lesions, vary in shape, and may bleed. These are the early symptoms of more typical brain cancer patients. In recent years, the level of diagnosis and treatment of brain tumors is constantly improving. Computed tomography (CT) and magnetic resonance imaging (MRI) and other imaging technologies have also been widely used in brain tumor detection, which is very effective for patients to check brain tumor, And the detection rate is relatively high. MRI has become an important part of modern imaging medicine. According to relevant clinical research, MRI is better than CT in the diagnosis and location of intracranial brain tumors, and can reach a 98% accuracy rate. MRI imaging technology has many advantages: people are more concerned about radiation, but this imaging technology will not cause harm to the human body; multiple parameters can be used for imaging, this imaging method can provide a wealth of useful information to facilitate Diagnosis is also more convenient and effective to study human metabolism and function. In addition, MRI imaging technology provides a wealth of anatomy of human soft tissue [3]. In medicine, MRI imaging technology is generally used to segment and detect brain tumors. Experts determine the existence and development of brain tumors by analyzing MR images. In this article, brain tumor slice images obtained by MRI imaging technology will also be used for experiments.
Because medical images are closely linked to practical applications and imaging medicine has also achieved success in clinical medicine, many countries have invested a lot of financial and human resources to research the processing of medical images and have achieved certain results. Image segmentation has become more and more important in image medicine. Image segmentation is an indispensable method when extracting quantitative information from image images. Segmented images can be used in many aspects, such as localization and diagnosis. Lesion tissue, quantitative analysis of tissue volume, learning of anatomical structure, etc., but the segmentation of medical images has not achieved particularly good results today, because medical images are more complex and diverse, and the tissues themselves have special differences, which leads to The formation of medical images is susceptible to noise, local body effects, and tissue movement. In addition, the structure and shape of the human anatomy are quite complicated, and there are differences between each person. This brings medical image segmentation. Great difficulty. It is difficult to obtain good results if traditional segmentation techniques are applied to medical images, such as a single segmentation technique or segmentation based on a single feature information. The clinical requirements for medical image segmentation are very high, with high rigor, and the speed of the algorithm is also required. The shorter the consumption time, the better.

II. RELATED WORK
In the basic research of image retrieval, Ranjith et al. [9] proposed a feature extraction method combining color, shape and texture features and using genetic algorithm to measure similarity. Pisapia et al. [10] proposed to retrieve images using the spatial data attributes of images. Ibragimov et al. [11] proposed a region-based image retrieval method using shapeadaptive discrete cosine transform. Kassahun et al. [12] proposed a feature fusion method for image retrieval based on a class of support vector machines. Cordova et al. [13] proposed a complementary semantic model for content-based image retrieval. Madubata et al. [14] proposed a feature extraction method using gray level co-occurrence matrix for image retrieval and segmentation. O'Donnell et al. [15] proposed a method for image retrieval using the ABIR algorithm. Gebru et al. [16] used local descriptor modeling as a contribution function, and then proposed a new multi-task strategy to improve the feature word bag.
Mires developed by the Chinese Academy of Sciences is a search tool with an image browsing function for crossplatform high-dimensional spatial indexing. It mainly adopts a similar matching technology, and the feature extraction mainly adopts three characteristics of shape, texture and color. It also effectively supports text-based image retrieval, and the search user can also evaluate the quality of the search results, and the system will also provide timely and effective feedback based on the evaluation of the search user [17]. In China, researchers from Shanghai Jiaotong University conducted a search test on the use of mutual information algorithms for ultrasound medical images. The computational complexity of this system is relatively low. Moreover, it extracts features from medical images, which breaks through the traditional retrieval ideas based on expressions and texts and has strong anti-interference. In addition, the image scalability, rotation, and position have little effect on the results of the search, and a retrieval framework for ultrasonic images has been initially formed, which has certain application significance [18].
Rabaglino et al. [19] proposed a feature extraction method that combines global features and SIFT features. Tseng et al. [20] proposed a method of using the local similarity of positional relations as the layout feature of images. Papp et al. [21] used the artificial fish swarm algorithm to adjust the feature weights using the user's feedback information. Goryawala et al. [22] combined the mixed features of grayscale features and texture features to derive potential topic frameworks and high-level semantic connections. Elton and Yalowich [23] proposed a cloud computing-based retrieval system using Brushlet transform and LBP algorithm. Beck et al. [24] analyzed and compared the main methods based on feature extraction such as color, texture, shape and spatial relationship, and proposed image retrieval using various features comprehensively.
The role of image segmentation in imaging medicine is growing. Image segmentation is an indispensable method when extracting quantitative information from image images. Moreover, the segmented image can be applied to many aspects, such as locating and diagnosing diseased tissue, quantitatively analyzing tissue volume, and learning anatomy. However, the segmentation of medical images has not achieved particularly good results today. The reason is that medical images are complex and diverse. Moreover, the organization itself has special differences, which leads to the influence of noise, local body effects, tissue movement, etc. when forming medical images. In addition, the structure and shape of the human anatomy are quite complicated, and there are differences among people, which brings great difficulties to the segmentation of medical images. Based on this, this study, based on the machine learning method, constructs a model of brain tumor feature monitoring points to improve the efficiency of brain tumor detection.

III. FEATURE EXTRACTION METHOD
Principal Component Analysis (PCA) is a classic feature extraction method, which is based on statistics. The essence is to project the sample data in the high-dimensional space into the low-dimensional space by linear transformation under the premise of representing the original data as well as possible, thereby extracting the main features of the data in the original space, and removing the correlation between the features, that is, removing the redundant components of the information. The PCA is built on the basis of the K-L expansion.
The K-L transform is a common orthogonal transform that is used to highlight differences and reduce correlation. Assuming X is an n-dimensional random variable, the weighted sum of n orthogonal basis vectors is used to represent x as: Among them, ϕ i is an orthogonal basis vector, α i is a weighting coefficient, and ϕ i satisfies: The matrix of equation (1) is expressed as: In the formula, α = (α 1 , α 2 , · · · , α n ) T , = (ϕ 1 , ϕ 2 , · · · , ϕ n ) are orthogonal matrices, i.e. T = I . Considering that is an orthogonal matrix, after two sides of equation (3) are pre-multiplied by T , the following equation can be obtained: We assume that the overall autocorrelation matrix for x is: After formula (3) is substituted into formula (5), the following formula is obtained.
If the components of the required vector α are not related to each other, the following formula should be satisfied.
Equation (7) is written in a rectangular form, After formula (8) is substituted into formula (6), the following formula can be obtained: We post-multiply the two sides of the above formula by . Then, since is an orthogonal matrix, the following formula is obtained.
That is: From equation (11), λ j is the eigenvalue of the autocorrelation matrix R, and ϕ j is the corresponding eigenvector. Since R is a real pair matrix, different eigenvalues are orthogonal to the corresponding eigenvectors. VOLUME 8, 2020

IV. PCA BASED FEATURE EXTRACTION
Assume that for an n-dimensional input sample x, there is a formula M-dimensional Y is obtained. Among them, n > m. If Y is a feature of X, that is, Y retains most of the feature information of x, then the feature extraction of PCA is the process of obtaining this transformation matrix A. It can be known from formula (11) that m-number eigenvectors constitute A, that is, A = [ϕ 1 , ϕ 2 , · · · , ϕ m ] can meet the requirements. Then, the problem is transformed into the selection of the eigenvectors, so that the matrix A of the constituents obtained by the formula (12) has the smallest square error of Y and x.
For equation (1), only m terms are taken, and the remaining n − m terms are replaced by constant b j , then the estimated value of x is Then the error between the estimated valuesx and x is Mean square error is It can be seen from equation (15) that when the mean square error ε 2 is minimized, b j should satisfy It can be obtained that the constant b j in equation (13) should be replaced by the expected value of the remaining n − m. A new coordinate system is set up, and the overall mean is taken as the origin of the new coordinate system, that is, Then, according to formula (17), formula (15) can be rewritten as Among them, λ j is the eigenvalue of the autocorrelation matrix R of x, and ϕ j is the eigenvector corresponding to λ j .
It can be known from equation (18) that the smaller the value of λ j , the smaller the mean square error. Therefore, α is defined as the energy retention rate after feature extraction, and it needs to satisfy the following formula after feature extraction.
The steps for extracting the entire PCA feature are as follows: (1) The mean vector of the mode population is taken as the origin, and the coordinate system is translated; (2) According to formula (5), the autocorrelation matrix R is obtained; (3) According to formula (11), the eigenvalue λ 1 , λ 2 , · · · , λ n of the autocorrelation matrix R and the corresponding eigenvector ϕ 1 , ϕ 2 , · · · , ϕ n are obtained; (4) The obtained eigenvalues are sorted. Then, the eigenvectors corresponding to the ordered eigenvalues of the first m numbers are selected to form a transformation moment.

V. IMPROVED FEATURE EXTRACTION METHOD
This article mentions the principle of feature extraction for automatic encoders. However, it is not sufficient for the output Z of the automatic encoder to maintain the reconstruction criteria of the information of the input X as much as possible. In particular, when the hidden layer has the same dimensions as the node of input X, a simple identity map can achieve perfect reconstruction. If there are no other constraints, SAE cannot get a useful feature extraction model from training.
Vincent et al. proposed and explored a different approachdenoising autoencoders. It differs from the representation Y of the constrained hidden layer, but changes the refactoring criteria for a more challenging target, that is the portion of the input X that is corrupted needs to be cleared or denoised by the reconstructed output Z. In this regard, the definition of ''a good representation'' is modified: A good representation is available from a corrupted input sample, and this is very useful for refactoring a clean output. This approach implies two basic ideas: (1) A high level of hidden layer representation should be stable and robust under corrupted sample input; (2) Performing a good deconvolution task requires obtaining a useful structure under the input distribution to extract features.
The denoising autoencoder is a variant of an autoencoder. First, by randomly mappingx ∼ q D (x |x ), the initial input X is converted to x. The random mapx ∼ q D (x |x ) is a process of mapping the original clean input x to the corrupted input x. In this article, masking noise is chosen, that is, for each clean input sample X, the value of the element inside is randomly limited to 0 by a fixed percentage. It can be thought of as setting the default values for those missing nodes or replacing them-the so-called common technique for handling missing values. All information about these masking nodes is thus removed from the particular input mode, and the denoising autoencoder can be seen as a tool trained to fill these manually introduced ''blanks''. Furthermore, forcing a node to zero means that they are completely ignored in the calculation of the next layer of neurons. After getting the mapped input x, x is used as the input to the autoencoder and input to the encoder. After that, the transformed activation value y is obtained by the formula (20): Among them, S is a nonlinear activation function, such as the commonly used sigmoid function.
θ = {W , b} is the parameter set, w is the weight of the encoder, and b is the offset of the presentation layer. After that, Y is input to a decoder, then the output z is obtained by the formula (22): Among them, θ = W , b is the parameter set, W is the weight of the decoder, and b is the offset of the reconstruction layer. At this time, Z is the reconstruction of the inputx after the damage. As shown in Fig. 1, it is a rough representation of the operation of the denoising autoencoder. Through a training set, θ and θ are trained to minimize the average reconstruction error loss, that is, to get the resulting output z as close as possible to the non-corrupted input X. The main difference here is that z is now a deterministic function ofx instead of x.
Or cross entropy error function: The parameters are randomly initialized and optimized by Random gradient descent method. The denoising autoencoder is to minimize the reconstruction loss between a clean x and the reconstructed z from y. Therefore, it raises the lower limit of the interaction information between a clean input x and express y. The difference is that y is now obtained by a corrupted input samplex by equation (20). Therefore, the denoising autoencoder forces a learning of a map that is smarter than an autoencoder: A feature extraction useful for denoising.

FIGURE 2. Manifold learning interpretation.
The denoising process maps a corrupted sample input to an undamaged output, which can give an intuitive geometric interpretation in the so-called manifold hypothesis. This manifold hypothesis indicates that the natural high-dimensional data set is close to a nonlinear low-dimensional manifold. As shown in Fig. 2, assuming that the training number (x) is concentrated close to a low-dimensional manifold, the damaged sample (·) obtained by the random mapping q D (x |x ) is generally far from the manifold. By learning p (x |x ), the model puts them back into the manifold (via the autoencoder g θ (f θ (·))) and the middle layer represent that y = f θ (x) can be interpreted as the coordinate system of point x on the manifold. During the denoising training process, the denoising autoencoder learns a random operator p (x |x ) and maps a corruptedx to a clean x.
A damaged sample is more likely to be far from the manifold than a clean sample. Therefore, the mapping learned by the random operator p (x |x ) tends to approach the point x of high probability from the point Z with a lower probability, which is on or near the manifold. It is worth noting that wheñ x is far from the manifold, p (x |x ) should learn to make more steps to reach the manifold. Successful denoising means that the random operator moves from a point farther away to a small area close to the manifold. Therefore, the denoising autoencoder can be seen as a way to define and learn a manifold. In particular, if the dimension limiting y is smaller than the dimension of x, then the intermediate representation layer y = f (x) may be viewed as a coordinate system of points on the manifold. It can generally be considered that y = f (x) is a representation of x, which is very suitable for capturing the main variables in the data. VOLUME 8, 2020 A stacking denoising autoencoder is used to initialize a deep learning network that is roughly the same as a traditional stacked autoencoder. The difference is that the stacking autoencoder uses the output of the L − 1 layer as the input of the L layer, while the stack denoising autoencoder uses the output of the L − 1 layer as a clean input, and randomly maps x ∼ q D (x |x ), that is,x x after masking noise, and is used as an input of the L layer. In addition to some initialization parameters of the autoencoder, the masking noise parameters of the denoising autoencoder are also independent, so the appropriate parameters can be selected for the masking noise of each denoising autoencoder. Once a stacked encoder is thus built, his top-level output representation can be used for an independent supervised learning algorithm and a deep neural network is generated for supervised learning. Moreover, the parameters of all layers can be fine-tuned using a stochastic gradient descent method.

VI. EXPERIMENT RESULTS
The data set used in this experiment is the GBM data set, GBM refers to pleomorphic glioblastoma, and each patient's brain tumor section includes brain tumor part, edema part and normal brain tissue part. The reason why GBM brain tumors are chosen is because it is the most common primary tumor in the central nervous system, accounting for 40% of brain tumor patients of all ages. The MR image slices for each patient in the data set ranged from 19 to 24, and each slice included an image of four channels and a corresponding marker image. In the experiment of this chapter, the brain tumor slice image of one patient was selected, and Fig.3 and Fig.4 are the images of the four channels corresponding to the eighth slice of the brain tumor of the B patient and the expert annotation image. In this article, the pixel samples extracted by patient B are used as training sets. When segmenting brain tumors, the pixels marked as brain tumors are positive samples, others are negative samples. Moreover, when dividing the edema, the same treatment is taken. Each SVM-based brain tumor segmentation is a problem of two classifications. The sample training LSSVM model is extracted, and then a brain tumor   slice of one patient (Life A) is selected for testing, and the classification result is expressed in the image to obtain the segmentation result. After that, post-processing is performed to obtain the final segmentation result. Fig. 5 and Fig.6 are the results of segmentation of the second slice in A, and the image corresponding to the second slice in A contains only edema and no contains brain tumor.
After the features extracted by Gabor wavelet and convolutional neural network are merged, each sample will get a new feature of 352 dimensions, which consumes a lot of time when sorting with LSSVM. In this article, the KECA method is used to reduce the dimension of the merged features. At the same time, the experimental results are used to compare and  analyze the influence of KECA dimension reduction on the classification results and whether it can meet its requirements. Fig. 7 is a Jaccard value obtained by dropping the different dimensions of the Gabor feature and the CNN feature of all sample points of the edema of the 6th slice brain tumor and the 12th slice in the A patient. Fig. 8 shows the finalized segmentation results obtained by reducing the dimensions of the merged features, classifying them with LSSVM, and then morphologically processing them. The first column is the expert labeling map of each slice in A, the second column is the brain tumor segmentation map obtained after dimension reduction, and the third column is the edema segmentation map obtained after dimension reduction. Moreover, the segmentation maps obtained in the Fig.9 are all segmentation maps obtained in the case of dropping to 200 dimensions.
After merging the features extracted by Gabor wavelet and convolutional neural network, each sample will get 352-dimensional new features, which will take a long time to perform classification processing with LSSVM. In this article, the KECA method will be used to reduce the dimensions of the fused features, and the experimental results will be compared to analyze the impact of KECA's dimensionality reduction on the classification result and whether it can meet its requirements. Fig. 10 shows the Jaccard values obtained by reducing the dimensionality of the Gabor features and  CNN features of all the sample points of the 6th slice brain tumor and the 12th slice edema in A.
It can be seen from Fig. 10 that when KECA is used to reduce the number of dimensions, the reduction of different dimensions has no significant effect on the segmentation results of brain tumors and edema, and the results obtained after reducing the dimensions are relatively stable, compared to KPCA. KPCA's segmentation results are very unstable when reducing the number of dimensions, and the Jaccard value of the segmentation results after dimensionality reduction also drops a lot, and there is a large fluctuation. The results obtained from different dimensions from 1 to 352 are very unstable. It can be seen that the loss of data information by the KECA algorithm in the process of dimensionality reduction is relatively small compared to the KPCA algorithm. From the experimental results, it is known that KECA works best when it is reduced to about 200, and can achieve the segmentation accuracy before dimensionality reduction. The least effective information was lost. Fig. 11 shows the final segmentation result obtained by performing dimensionality reduction on the fused features, classifying them with LSSVM, and then using morphological processing. The first column is the expert labeling map of each slice in A, and the second column of segmentation map FIGURE 11. Segmentation results of brain tumors in section A after dimensionality reduction. for brain tumors obtained after dimensionality reduction. The third column is the edema segmentation map obtained after dimensionality reduction.

VII. DISCUSSION
As can be seen from Fig. 3, the gray part in the middle of the image marked by the expert is a tumor, the white part is edema, and the gray part outside is the normal tissue of the brain. Extract training samples from the patient's brain tumor slices: Firstly, four pixel points corresponding to four channels are taken out as a sample according to the mark image, and the sample corresponding to the pixel point marked as a brain tumor is taken as a positive sample of brain tumor, and other pixels are negative samples with respect to the brain tumor. According to the mark of the edema portion in the marked image, the pixels of the four channels representing the edema are presented as the edema sample, and the corresponding sample labeled as the edema is the edema positive sample, and the other relative edema is expressed as the negative sample.
As shown in Fig. 5 and Fig. 6, there is a lot of noise in the segmentation result graph. In this article, the image processing is based on pixel points, and when LSSVM classifies each pixel, the error condition will appear. Therefore, it needs to be processed morphologically. Moreover, this article uses the open operation in morphology to take the process, and the structural elements use the disc structure. The open operation is to first perform the etching operation on the segmented image, and then perform the expansion operation, which can eliminate the small objects in the image, eliminate the burrs, smooth the edges of the large objects, and eliminate the noise. The processing results are different depending on the radius of the disc, and the numbers in parentheses in the Fig. represent the radius.
It can be seen from Fig. 7 that the effect of reducing the different dimensions in the dimension reduction treatment with KECA on the segmentation results of brain tumors and edema is not particularly large, and the results obtained after reducing the different dimensions are relatively stable. Compared with KPCA, KPCA's segmentation results are very unstable when the different dimensions are reduced, the Jaccard value of the segmentation results after the dimension reduction is also much lower, and there are large fluctuations, and the results obtained from 1D to 352 dimensions are very unstable. It can be seen that the loss of data information in the process of dimension reduction by the KECA algorithm is small compared to the KPCA algorithm. In addition, from the experimental results, it can be seen that KECA has the best effect when it is reduced to about 200, and can achieve the segmentation accuracy before dimension reduction, and the loss of effective information is the least.
It can be seen from Fig. 8, Fig. 9 and Fig. 10 that, except that the bold Jaccard value is higher than before the dimension reduction, the others are not much different from the predimensionality reduction, and the segmentation accuracy is not greatly reduced. Because KECA retains the main valid information of the original data to the greatest extent according to the size of the entropy, it can better retain the characteristic information of the original data. Therefore, even if the dimension of the drop in the dimension reduction process is low, it can also produce better results. In addition, the results of the post-dimension reduction in individual cases are equal to or greater than those before the dimension reduction. Although the feature information after feature fusion will be more than that before the fusion, it is also possible to add redundant invalid information, so that the segmentation precision is reduced, and more compact information can be obtained after dimension reduction, the included features can better represent the data information, and the segmentation precision can be improved. Moreover, after dimension reduction, it is obvious that the computational complexity can be reduced, thereby reducing the computation time of subsequent classification. Finally, the experiment proves that using KECA for dimensionality reduction achieves a balance between time and precision.

VIII. CONCLUSION
Based on the machine learning method, this study constructs a model for monitoring the characteristics of brain tumors and strives to improve the efficiency of brain tumor detection. In the study, the sample data in the high-dimensional space is projected into the low-dimensional space by linear transformation, thereby extracting the main features of the data in the original space and removing the correlation between the features. The process of denoising refers to mapping a corrupted sample input to an undamaged output, which gives an intuitive geometric interpretation of the so-called manifold hypothesis. This manifold hypothesis indicates that the natural high-dimensional data set is close to a nonlinear lowdimensional manifold. In addition, the data set used in this experiment is the GBM data set, GBM refers to pleomorphic glioblastoma, and each patient's brain tumor section includes brain tumor part, edema part and normal brain tissue part. In summary, the experimental results show that the method proposed in this study has a certain effect in brain tumor detection, which can provide a theoretical reference for subsequent clinical research. He is currently a famous TCM Expert and has a long career in curing mental, centrum cerebrovascular, and digestive diseases. He is also an Academic Leader of the Nei Jing in State Administration of TCM. His research interests include TCM fundamental theories and Huang Di Nei Jing.