Chemical exchange saturation transfer image anomaly detection method

Abstract

The invention discloses a chemical exchange saturation transfer image anomaly detection method, which comprises the following steps: preprocessing CEST data obtained by scanning to obtain a spectrum data set of all pixels, and finally dividing the spectrum data into a training set and a test set; the method comprises the following steps: constructing an auto-encoder model Mamb-AE based on a Mamb network; constructing a joint loss function of a feature space and a data space, wherein the joint loss function is used for Mamba-AE training; inputting the training set into the Mamb-AE, and training the Mamb-AE by using a joint loss function; an abnormal score calculation mechanism is constructed and used for judging the possibility that the pixel belongs to the focus; and inputting the divided test set into the trained Mamba-AE, outputting abnormal scores pixel by pixel based on an abnormal score calculation mechanism, and returning all the abnormal scores to each corresponding pixel to form an abnormal image as a focus detection result.

Description

Technical Field

[0001] The invention belongs to the technical field of medical magnetic resonance image analysis, and in particular relates to a method for detecting abnormalities in chemical exchange saturation transfer images. Background Art

[0002] Chemical exchange saturation transfer (CEST) magnetic resonance imaging (MRI) is an emerging molecular imaging technology that can specifically detect low concentrations of endogenous metabolites, such as glucose, glutamate, creatinine, etc., and use them as non-invasive biomarkers. Due to this feature, many studies have shown that CEST imaging has the ability to detect a variety of diseases, including tumors, Alzheimer's disease and stroke, so this technology has broad prospects for clinical application. However, the CEST scanning process is easily interfered by a variety of accompanying signals that are insensitive to lesion manifestations, such as direct water saturation (DS) effect and magnetization transfer effect (MT). These effects dilute key diagnostic information and severely weaken the contrast of lesion tissue in the image, posing challenges to clinical diagnosis.

[0003] It is very difficult to manually detect lesions from disturbed images, so it is a feasible means to use computer-aided diagnosis technology for detection. In recent years, deep neural networks have become a popular medical image analysis tool that can help researchers extract pathological information from mixed signals. Although supervised learning methods have performed relatively stably in disease detection and segmentation tasks, this method relies on high-quality labeled data. The acquisition and annotation of clinical pathological images is a challenging task, and it is also difficult to summarize the different abnormal manifestations of various diseases when constructing data sets. Therefore, compared with supervised methods, unsupervised anomaly detection without annotated pathological data has received increasing attention. This method aims to identify abnormal information by learning the patterns of normal samples. It usually uses a generative model to reconstruct the input image and judge whether the data is abnormal based on the difference between the original input and the reconstructed data, that is, the size of the reconstruction error. Generally speaking, lesion data will show significantly higher reconstruction errors than normal data. However, traditional methods are not suitable for lesion detection in CEST images. The main reason is that each pixel is composed of long-range spectra, which contain pathological response signals of multiple metabolites. Therefore, the detection model needs to have the ability to model the spectrum sequence. However, the traditional convolutional neural network (CNN) is limited by its receptive field and lacks the ability to model long distances; the recurrent neural network (RNN) and long short-term memory network (LSTM) models have difficulty in providing a comprehensive description of the sequence context by learning the sequence point by point; the Transformer model has limited performance due to the quadratic computational complexity brought by its self-attention mechanism. Therefore, it is of great value to develop an efficient and high-precision lesion detection method based on the characteristics of CEST MRI. Summary of the invention

[0004] The purpose of the present invention is to provide a chemical exchange saturation transfer image anomaly detection method to solve the computational complexity problem existing in the existing method so as to improve the detection accuracy and robustness.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is: a method for detecting abnormality of chemical exchange saturation transfer images, which is specifically implemented according to the following steps:

[0006] Step 1: pre-process the scanned CEST data to obtain a spectrum data set of all pixels, and finally divide the spectrum data into a training set and a test set;

[0007] Step 2, construct the autoencoder model Mamba-AE based on the Mamba network;

[0008] Step 3: construct a joint loss function of feature space and data space for Mamba-AE training;

[0009] Step 4, input the training set into Mamba-AE, and train Mamba-AE using the joint loss function;

[0010] Step 5: construct an abnormality score calculation mechanism to judge the possibility that the pixel belongs to a lesion;

[0011] In step 6, the divided test set is input into the trained Mamba-AE, and the anomaly score is output pixel by pixel based on the anomaly score calculation mechanism, and all anomaly scores are returned to each corresponding pixel to form an abnormal image as the lesion detection result.

[0012] As a preferred technical solution of the present invention, the step 1 is specifically:

[0013] Step 1.1, using the signal before applying the RF saturation pulse in the CEST process to normalize the original CEST data, and map it to the (0,1) interval;

[0014] Step 1.2, converting the dimension of the data set. The original data set dimension is a four-dimensional format of sample size, image length, image width, and spectrum frequency. In order to adapt to the characteristics of the method of the present invention, it is converted into a two-dimensional data format of sample size × image length × image width, spectrum frequency;

[0015] Step 1.3, extract part of the spectrum from the normal sample in the converted data and divide it into the training set, and divide the rest of the spectrum into the test set.

[0016] As a preferred technical solution of the present invention, the step 2 is specifically:

[0017] Step 2.1, construct the Mamba module as the main unit of the Mamba-AE model, which extracts features by establishing a mapping between sequence data;

[0018] In step 2.2, multiple encoder layers and decoder layers are constructed, each of which is composed of multiple Mamba modules of different scales.

[0019] As a preferred technical solution of the present invention, the step 2.1 is specifically:

[0020] Step 2.1.1, construct a state space model (SSM) with a selection mechanism, which is used for the state evolution of the spectrum sequence at multiple time steps, predicts its hidden state at the next time step based on the input, and finally outputs the predicted sequence. The model can be described as an ordinary differential state equation; construct a selection mechanism to convert the parameters of the equation into a function of the input through a linear layer;

[0021] Step 2.1.2, construct the Mamba module, whose main body includes a root mean square error normalization layer, a dual-branch path structure, and a residual connection structure; one of the dual-branch paths includes a linear projection layer, a convolution layer, an activation function layer, and the above-mentioned SSM layer in sequence, and the other path includes a linear projection layer and an activation function layer. The outputs of the two paths are fused by point-by-point multiplication, and the fused features are input through the residual connection.

[0022] As a preferred technical solution of the present invention, the step 2.2 is specifically:

[0023] Step 2.2.1, construct a multi-layer encoder to extract the features of the input spectrum sequence. Each encoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each encoder layer, the output sequence size is downsampled to half of the input.

[0024] Step 2.2.2, construct a multi-layer decoder to reconstruct the input spectral features into the original sequence. Each decoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each decoder layer, the output sequence size is upsampled to twice the input.

[0025] As a preferred technical solution of the present invention, in step 3, specifically:

[0026] Step 3.1, calculate the Huber loss between the original sequence and the generated sequence, construct the data reconstruction loss function, and adaptively select the specific calculation strategy of the loss by comparing the difference between the preset threshold and the original-reconstructed sequence difference;

[0027] Step 3.2, construct the feature reconstruction loss function. For the multi-layer encoder-decoder structure, except for the outermost layer that directly processes the input and output data, the cosine similarity between the input features of each encoder layer and the output features of the corresponding decoder layer is calculated pixel by pixel at different scales as the feature reconstruction loss.

[0028] In step 3.3, weights are assigned to the data reconstruction loss function and the feature reconstruction loss function respectively, and the weighted sum is used as the total loss function for Mamba-AE model training.

[0029] As a preferred technical solution of the present invention, in step 5, an abnormality score calculation mechanism is constructed to calculate a corresponding abnormality score according to the reconstructed spectrum of each pixel, which is used to judge the possibility that the pixel belongs to a lesion.

[0030] As a preferred technical solution of the present invention, in step 5, specifically:

[0031] Step 5.1, construct an anomaly score calculation method for reconstructed sequence data, and use the mean square error to calculate the difference between the reconstructed sequence output by the model and the original input sequence as the data anomaly score;

[0032] Step 5.2: Construct a method for calculating the anomaly score of the reconstructed features, extract the input features of each layer of encoder and the output features of each layer of decoder during the test, calculate the cosine similarity of the features between the corresponding layers, and add the results one by one as the overall feature anomaly score;

[0033] In step 5.3, add the data anomaly score and the feature anomaly score to obtain the final anomaly score.

[0034] The beneficial effects of the present invention are as follows: the chemical exchange saturation transfer imaging anomaly detection method of the present invention is used for unsupervised detection of CESTMRI lesions. The model has a strong ability to capture long-range dependencies and effectively models CEST spectrum sequences with low computational complexity; in addition, the present invention also designs a joint loss function, which combines the reconstruction loss based on the paired features between the encoder and the decoder and the reconstruction loss in the data space. The design enables the model to simultaneously learn the semantic patterns of normal samples and physiological tissue structures. The present invention has high clinical application value, can help doctors identify and locate lesions in the early stages of the disease, significantly improve the efficiency and accuracy of clinical diagnosis, and improve the prognosis of treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0036] Figure 1 An overall flow chart of an unsupervised lesion detection method using chemical exchange saturation transfer imaging provided by an embodiment of the present invention;

[0037] Figure 2 A structural diagram of the Mamba module provided by an embodiment of the present invention;

[0038] Figure 3 A structural diagram of the autoencoder and feature-data joint anomaly assessment mechanism provided in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram comparing the lesion detection results of the method provided by the present invention and the existing method in the rat stroke data set 2 hours, 6 hours and 24 hours after ischemia. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] like Figure 1 As shown, a method for detecting abnormality of chemical exchange saturation transfer images of the present invention is specifically implemented according to the following steps:

[0043] Step 1: pre-process the scanned CEST data to obtain a spectrum data set of all pixels, and finally divide the spectrum data into a training set and a test set;

[0044] Step 2, construct the autoencoder model Mamba-AE based on the Mamba network;

[0045] Step 3: construct a joint loss function of feature space and data space for Mamba-AE training;

[0046] Step 4, input the training set into Mamba-AE, and train Mamba-AE using the joint loss function;

[0047] Step 5, construct an abnormality score calculation mechanism to calculate the corresponding abnormality score according to the reconstructed spectrum of each pixel, which is used to judge the possibility that the pixel belongs to the lesion;

[0048] In step 6, the divided test set is input into the trained Mamba-AE, and the anomaly score calculation mechanism established in step 5 outputs the anomaly score pixel by pixel, and all anomaly scores are returned to each corresponding pixel to form an abnormal image as the lesion detection result.

[0049] The present invention takes unsupervised lesion detection in CEST MRI as the research object. Aiming at the shortcomings of traditional methods in detection accuracy and computational complexity, an autoencoder model Mamba-AE based on Mamba network is designed. The model reconstructs the input CEST spectrum to evaluate its abnormality, uses multiple Mamba modules to capture long-range dependencies in the spectrum to learn the signal pattern of normal tissue, and introduces reconstruction losses in the data space and potential feature space to jointly constrain the learning process of the model. The present invention can achieve accurate and robust pixel-by-pixel lesion detection.

[0050] Example 2

[0051] Different from Example 1, in the chemical exchange saturation transfer image abnormality detection method of Example 2 of the present invention, step 1 is specifically:

[0052] Step 1.1, using the signal before applying the RF saturation pulse in the CEST process to normalize the original CEST data, and map it to the (0, 1) interval;

[0053] Step 1.2, converting the dimension of the data set. The original data set dimension is a four-dimensional format of sample size, image length, image width, and spectrum frequency. In order to adapt to the characteristics of the method of the present invention, it is converted into a two-dimensional data format of sample size × image length × image width, spectrum frequency;

[0054] Step 1.3, extract part of the spectrum from the normal sample in the converted data and divide it into the training set, and divide the rest of the spectrum into the test set.

[0055] CEST MRI image data is collected and preprocessed to adapt it to the data format requirements of the method of the present invention. In the embodiment of the present invention, the data set used is from ischemic stroke rats, and the middle artery ischemic occlusion model (MCAO) is prepared by thread embolism. The scanning parameters are set to the main magnetic field strength of 7 Tesla, the saturated radio frequency pulse intensity of 1 μTelsa, and the scanning frequency range is -10ppm to 10ppm. The scans are performed at four time points before modeling, 2 hours, 6 hours, and 24 hours after modeling.

[0056] For all scanned CEST data S sat , the signal S0 before applying the RF saturation pulse during the scanning process is used for normalization to obtain the normalized data S, which is expressed as:

[0057] s=S sat / S0;

[0058] Then, the dimension of the data set is transformed. For the normalized data set D = {S1, S2, ..., S N}, which contains N groups of data, each of which has a dimension of (H, W, f), namely the length, width and frequency of the image, and is converted into a set of spectrum sequences with a dimension of (HW, f).

[0059] In this embodiment of the present invention, a data set containing 603 frequencies and a total of 9934 pixels is obtained, of which 3739 are from scans of normal samples and are divided into a training data set, and the remaining data are from scans of samples of three different pathological stages and are divided into a test data set.

[0060] Example 3

[0061] Different from Example 2, in Example 3, in the chemical exchange saturation transfer image abnormality detection method of the present invention, the specific process of step 2 is:

[0062] Step 2.1, construct the Mamba module as the main unit of the Mamba-AE model. This module extracts features by establishing a mapping between sequence data, specifically:

[0063] Step 2.1.1, construct a state space model (SSM) with a selection mechanism, which is used for the state evolution of the spectrum sequence at multiple time steps, predicts its hidden state at the next time step based on the input, and finally outputs the predicted sequence. The model can be described as an ordinary differential state equation; construct a selection mechanism to convert the parameters of the equation into a function of the input through a linear layer;

[0064] Step 2.1.2, construct the Mamba module, whose main body includes a root mean square error normalization layer, a dual-branch path structure, and a residual connection structure; one of the dual-branch paths includes a linear projection layer, a convolution layer, an activation function layer, and the above-mentioned SSM layer in sequence, and the other path includes a linear projection layer and an activation function layer. The outputs of the two paths are fused by point-by-point multiplication, and the fused features are input through the residual connection;

[0065] Step 2.2, construct multiple encoder layers and decoder layers, each of which is composed of multiple Mamba modules of different scales, specifically:

[0066] Step 2.2.1, construct a multi-layer encoder to extract the features of the input spectrum sequence. Each encoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each encoder layer, the output sequence size is downsampled to half of the input.

[0067] Step 2.2.2, construct a multi-layer decoder to reconstruct the input spectral features into the original sequence. Each decoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each decoder layer, the output sequence size is upsampled to twice the input.

[0068] An autoencoder model Mamba-AE based on the Mamba network is constructed. Mamba-AE includes an encoder and a decoder composed of the Mamba module as the backbone. Figure 2 The structure diagram of the Mamba module provided by the embodiment of the present invention is shown. The core of the module is a state space model (SSM) with a selection mechanism. In addition, it mainly includes a dual-branch path structure and a residual connection structure, specifically including:

[0069] SSM is used to establish the state evolution of a sequence at multiple time steps, predict the hidden state of the next time step based on the input, and finally output the target sequence. The model can be described by an ordinary differential equation:

[0070]

[0071] y t =Ch t +Dx t ;

[0072] Among them, A, B, C, and D are the coefficient matrices of the equation system, which are used to reflect the change of state over time, the influence of control input on state, the output generated based on the current state, and the direct influence of input on output; in order to give the model resolution invariance and ensure correct normalization, the time scale parameter Δ is introduced to discretize the continuous time system and convert the continuous parameters A and B into discrete parameters and In addition, in order to focus on or ignore specific inputs like the attention mechanism, a selection mechanism is added to adaptively adjust the values ​​of matrices B, C and time step Δ according to different input data. This is achieved by designing these parameters as functions of the input, and the expression is:

[0073] s B (x)=Linear N (x);

[0074] s C (x)=Linear N (x);

[0075] s Δ (x)=Linear D (x);

[0076] τ Δ =softplus;

[0077] The dual-branch path structure combines the input evolved by SSM with the original input for feature fusion. The input first passes through a layer of root mean square error normalization layer. One path specifically includes a linear projection layer, a 1×1 convolution layer, a SiLU activation function layer, and an SSM module; the other path includes a linear projection layer and a SiLU activation function layer; the inputs of the two paths are fused and output by point-by-point multiplication. This process can be expressed as

[0078] Shown as:

[0079] f′ i-1 =SSM(Silu(Conv((Linear(RN(f i-1 ))))))⊙Silu(Linear(RN(f i-1 )));

[0080] Among them, f i-1 and f′ i-1 They represent the input and output of the dual-branch path in the i-th Mamba module respectively;

[0081] The residual connection merges and normalizes the input into the Mamba module and its output after the dual-branch path structure. Specifically, the fusion is performed by addition. The process can be expressed as follows:

[0082] f i =f i-1 +Linear(f′ i-1 );

[0083] Among them, f i Represents the output of the ith Mamba module. If there is a connected next Mamba module, it will also serve as its input;

[0084] Figure 3 The structure diagram of the autoencoder based on the Mamba module is shown. Both the encoder and decoder models are based on multiple stacked Mamba modules. The specific structure is as follows:

[0085] Encoder: The encoder is used to extract and compress sequence features. In the embodiment of the present invention, for each encoder layer, four Mamba modules are stacked. Except for the Mamba module, the data is output after a layer of root mean square error normalization layer and a layer of linear projection layer. The output dimension of the last linear projection layer is one-half of the input dimension, so that the sequence length is halved. The input and output of each encoder can be expressed as follows:

[0086] f output =Linear(RN(MB4(MB3(MB2(MB1(f input ))))));

[0087] Decoder: The decoder is used to reconstruct and restore the sequence data. Similar to the encoder, in the embodiment of the present invention, for each decoder layer, four Mamba modules are stacked. Except for the Mamba module, the data is output after a layer of root mean square error normalization layer and a layer of linear projection layer. The output dimension of the last linear projection layer is twice the input dimension, so that the sequence length is restored to the original size step by step. The input and output expressions of each decoder are the same as those of the encoder.

[0088] Example 4

[0089] Different from Example 3, in Example 4 of the present invention, in the chemical exchange saturation transfer image abnormality detection method, the specific process of step 3 is:

[0090] Construct loss function for Mamb a -The joint loss function of AE model training consists of two parts: the reconstruction loss calculated using the paired features between the encoder and decoder and the reconstruction loss calculated using the input data and output data. Specifically, it includes:

[0091] Data space reconstruction loss: In the embodiment of the present invention, the generation model is constrained by measuring the distance between the original sequence and the generated sequence, and the Huber loss is selected as the calculation method. The calculation strategy is adaptively selected based on the comparison between the difference between the reconstructed value and the actual value and the preset threshold, so as to effectively process non-Gaussian noise while maintaining efficiency. The expression is:

[0092]

[0093] in, They represent the original data value and the predicted value respectively. The δ parameter is used as a weight to select the calculation formula. When the prediction error is less than δ, the mean square error (MSE) is used. Conversely, when the prediction error exceeds δ, a linear error approximating L1 is used. In the embodiment of the present invention, δ=1.

[0094] Feature space reconstruction loss: For the four-layer structure adopted in the embodiment of the present invention, in addition to the outermost layer directly processing the input and output data, the input features of the k-th layer encoder and the corresponding output features of the k-th layer decoder are used at different scales. By calculating the cosine similarity, the distance between the features is measured pixel by pixel as the feature reconstruction loss. The loss can be expressed as follows:

[0095]

[0096] in, is the feature of the k-th layer encoder / decoder, ||·|| represents the L2 norm;

[0097] Joint loss function: Based on the above two loss functions, the overall loss of optimizing the entire model can be expressed as:

[0098]

[0099] In the embodiment of the present invention, the weights α and β are set to 1 and 0.4 respectively.

[0100] Example 5

[0101] Different from Example 4, in the chemical exchange saturation transfer image abnormality detection method of Example 2 of the present invention, the specific process of step 4 is:

[0102] Mamba-AE model training. In the embodiment of the present invention, the model is implemented using the PyTorch framework, and the model parameters are optimized using the Adam optimizer. The learning rate is set to 0.0005, and other parameters include β1=0.5, β2=0.999. The model is trained for 200 rounds, and the batch size is 64. 。

[0103] Example 6

[0104] Different from Example 5, in the method for detecting abnormality of chemical exchange saturation transfer images of the present invention, step 5 is specifically as follows:

[0105] Step 5.1, construct an anomaly score calculation method for reconstructed sequence data, and use the mean square error to calculate the difference between the reconstructed sequence output by the model and the original input sequence as the data anomaly score;

[0106] Step 5.2, construct a method for calculating the anomaly score of the reconstructed features, extract the input features of each layer of encoder and the output features of each layer of decoder during the test, calculate the cosine similarity of the features between the corresponding layers, and add the results one by one as the overall feature anomaly score;

[0107] In step 5.3, add the data anomaly score and the feature anomaly score to obtain the final anomaly score.

[0108] The anomaly score calculation method is constructed, which is similar to the above-mentioned method of calculating the loss. The anomaly score of a single pixel consists of feature-level and data-level anomaly scores. The calculation method of the feature anomaly score is consistent with the calculation of the loss, which is achieved by evaluating the cosine similarity of the input features of the k-th layer encoder and the corresponding output features of the k-th layer decoder; while the data anomaly score is calculated using the MSE metric. The calculation method is as follows:

[0109]

[0110] Among them, a pair (h, w) is used to locate the position of a single pixel in the image. is the calculated anomaly score, the larger Values ​​represent higher likelihood of lesion.

[0111] Step 6 is as follows: Input the test set into the trained Mamba-AE model for testing. For each pixel’s spectrum y(h, w), an anomaly score is output using the above calculation method. All from the same image Reconstruct the abnormal image according to its corresponding position (h, w) And generate a visual heat map.

[0112] The above description shows and describes several preferred embodiments of the invention, but as mentioned above, it should be understood that the invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention concept described herein through the above teachings or the technology or knowledge of the relevant field. Changes and modifications made by those skilled in the art should be within the scope of protection of the claims attached to the invention without departing from the spirit and scope of the invention.

Claims

1. A method for detecting abnormality of chemical exchange saturation transfer images, characterized in that: include: Step 1: pre-process the scanned CEST data to obtain a spectrum data set of all pixels, and finally divide the spectrum data into a training set and a test set; Step 2, construct the autoencoder model Mamba-AE based on the Mamba network; Step 3: construct a joint loss function of feature space and data space for Mamba-AE training; Step 4, input the training set into Mamba-AE, and train Mamba-AE using the joint loss function; Step 5: construct an abnormality score calculation mechanism to judge the possibility that the pixel belongs to a lesion; In step 6, the divided test set is input into the trained Mamba-AE, and the anomaly score is output pixel by pixel based on the anomaly score calculation mechanism, and all anomaly scores are returned to each corresponding pixel to form an abnormal image as the lesion detection result.

2. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 1, characterized in that: The step 1 is specifically: Step 1.1, using the signal before applying the RF saturation pulse in the CEST process to normalize the original CEST data, and map it to the (0,1) interval; Step 1.2, converting the dimension of the data set. The original data set dimension is a four-dimensional format of sample size, image length, image width, and spectrum frequency. In order to adapt to the characteristics of the method of the present invention, it is converted into a two-dimensional data format of sample size × image length × image width, spectrum frequency; Step 1.3, extract part of the spectrum from the normal sample in the converted data and divide it into the training set, and divide the rest of the spectrum into the test set.

3. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 2, characterized in that: The step 2 is specifically as follows: Step 2.1, construct the Mamba module as the main unit of the Mamba-AE model, which extracts features by establishing a mapping between sequence data; In step 2.2, multiple encoder layers and decoder layers are constructed, each of which is composed of multiple Mamba modules of different scales.

4. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 3, characterized in that: The step 2.1 is specifically as follows: Step 2.1.1, construct a state space model (SSM) with a selection mechanism, which is used for the state evolution of the spectrum sequence at multiple time steps, predicts its hidden state at the next time step based on the input, and finally outputs the predicted sequence. The model can be described as an ordinary differential state equation; construct a selection mechanism to convert the parameters of the equation into a function of the input through a linear layer; Step 2.1.2, construct the Mamba module, whose main body includes a root mean square error normalization layer, a dual-branch path structure, and a residual connection structure; one of the dual-branch paths includes a linear projection layer, a convolution layer, an activation function layer, and the above-mentioned SSM layer in sequence, and the other path includes a linear projection layer and an activation function layer. The outputs of the two paths are fused by point-by-point multiplication, and the fused features are input through the residual connection.

5. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 4, characterized in that: The step 2.2 is specifically as follows: Step 2.2.1, construct a multi-layer encoder to extract the features of the input spectrum sequence. Each encoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each encoder layer, the output sequence size is downsampled to half of the input. Step 2.2.2, construct a multi-layer decoder to reconstruct the input spectral features into the original sequence. Each decoder layer includes multiple stacked Mamba modules, a root mean square error normalization layer, and a linear projection layer to change the size of the output sequence. After each decoder layer, the output sequence size is upsampled to twice the input.

6. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 5, characterized in that: In the step 3, specifically: Step 3.1, calculate the Huber loss between the original sequence and the generated sequence, construct the data reconstruction loss function, and adaptively select the specific calculation strategy of the loss by comparing the difference between the preset threshold and the original-reconstructed sequence difference; Step 3.2, construct the feature reconstruction loss function. For the multi-layer encoder-decoder structure, except for the outermost layer that directly processes the input and output data, the cosine similarity between the input features of each encoder layer and the output features of the corresponding decoder layer is calculated pixel by pixel at different scales as the feature reconstruction loss. In step 3.3, weights are assigned to the data reconstruction loss function and the feature reconstruction loss function respectively, and the weighted sum is used as the total loss function for Mamba-AE model training.

7. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 6, characterized in that: In step 5, an abnormality score calculation mechanism is constructed to calculate a corresponding abnormality score according to the reconstructed spectrum of each pixel, which is used to judge the possibility that the pixel belongs to a lesion.

8. The method for detecting abnormality of chemical exchange saturation transfer images according to claim 7, characterized in that: In step 5, specifically: Step 5.1, construct an anomaly score calculation method for reconstructed sequence data, and use the mean square error to calculate the difference between the reconstructed sequence output by the model and the original input sequence as the data anomaly score; Step 5.2, construct a method for calculating the anomaly score of the reconstructed features, extract the input features of each layer of encoder and the output features of each layer of decoder during the test, calculate the cosine similarity of the features between the corresponding layers, and add the results one by one as the overall feature anomaly score; In step 5.3, add the data anomaly score and the feature anomaly score to obtain the final anomaly score.

Images