Intelligent cerebral infarction MRI image lesion segmentation method based on deep learning
By selecting the proportion of gray values and irrelevance indices of MRI sequence images, a diverse sample set was constructed, which solved the problem of decreased model generalization ability caused by training sample redundancy and improved the accuracy and robustness of cerebral infarction lesion segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI TENTH PEOPLES HOSPITAL
- Filing Date
- 2026-01-28
- Publication Date
- 2026-06-26
AI Technical Summary
In existing neural network-based methods for segmenting lesions in MRI images of cerebral infarction, the high redundancy of training samples leads to a decrease in the model's generalization ability and affects the accuracy of lesion segmentation results.
By analyzing the grayscale value distribution of MRI sequence images, image groups with abnormal conditions were screened out, irrelevance index and difference were calculated, highly overlapping image groups were removed, and an effective sample set was constructed for training the neural network.
It enhances the model's generalization ability, improves the robustness of cerebral infarction lesion segmentation, avoids similarity misjudgment caused by redundant sequences, and improves the accuracy of lesion segmentation.
Smart Images
Figure CN121600265B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, specifically to a method for intelligent segmentation of lesions in MRI images of cerebral infarction based on deep learning. Background Technology
[0002] Cerebral infarction is a disease caused by the interruption of blood supply to the brain, leading to ischemia, hypoxia, and necrosis of brain tissue. MRI (Magnetic Resonance Imaging) is the gold standard imaging modality for diagnosing and evaluating cerebral infarction. It can clearly display brain tissue structure, especially DWI (Diffusion-Weighted Imaging) sequences, which can show high-signal lesions in the early stages of infarction. In MRI, the brightness (signal intensity) of each pixel reflects the physicochemical state of water molecules within the tissue at that location (such as content, degrees of freedom, and surrounding environment). The pathological process of cerebral infarction alters these states, thereby changing the MRI signal, thus enabling early diagnostic assessment through MRI images.
[0003] Existing neural network-based intelligent segmentation methods rely on full-sequence MRI data of cerebral infarction. When training the neural network, redundant sequences are present, which can easily lead to problems such as high similarity between samples, model overfitting, and decreased generalization ability, thus affecting the accuracy of lesion segmentation results. Summary of the Invention
[0004] To address the technical problem that existing methods suffer from high redundancy in training samples when training neural networks, leading to decreased accuracy in lesion segmentation, this invention aims to provide a deep learning-based intelligent lesion segmentation method for MRI images of cerebral infarction. The specific technical solution adopted is as follows:
[0005] Acquire an image set of each MRI sequence for cerebral infarction symptoms, wherein the image set contains several image groups;
[0006] Based on the distribution of grayscale values in each image within the image set of each MRI sequence, image groups with anomalies are screened to obtain the target group for each MRI sequence;
[0007] Based on the degree of difference between abnormalities in the same MRI sequence for each two target groups, and the degree of difference between each two target groups in each MRI sequence and other types of MRI sequences, the irrelevance index of each two target groups in the same MRI sequence is obtained.
[0008] The similarity index between each pair of target groups is obtained based on the irrelevance index and the difference index corresponding to each pair of target groups; based on the similarity index between each target group and other target groups, highly overlapping image groups are eliminated to obtain an effective sample set, which is used to train the intelligent segmentation neural network.
[0009] Preferably, the step of filtering image groups with abnormalities based on the grayscale value distribution of each image in the image set of each MRI sequence to obtain the target group for each MRI sequence specifically includes:
[0010] Based on the grayscale value of each pixel in each image of each MRI sequence, the pixels of all images in the same MRI sequence are clustered to obtain pixel clusters for each MRI sequence.
[0011] Based on the distribution of gray values in pixel clusters under each MRI sequence, the gray level of each pixel in each image in the image set of each MRI sequence is obtained;
[0012] The degree of abnormality in each image group of each MRI sequence is obtained by considering the proportion of pixels at different gray levels in each image group of each MRI sequence.
[0013] The image group corresponding to the abnormality level being greater than or equal to the preset abnormality threshold is designated as the target group for the corresponding MRI sequence.
[0014] Preferably, the step of obtaining the gray level of each pixel in each image of each MRI sequence image set based on the gray value distribution in pixel clusters under each MRI sequence specifically includes:
[0015] For any MRI sequence, all pixel clusters are arranged in ascending order of gray value. The gray level of each pixel in each image of the MRI sequence is determined by the arrangement number of the pixel cluster to which each pixel belongs.
[0016] Preferably, the step of determining the degree of abnormality in each image group of each MRI sequence based on the proportion of pixels at different gray levels in each image group of each MRI sequence specifically includes:
[0017] For any MRI sequence, the percentage of pixels at each gray level in each image is obtained as the quantitative feature coefficient of each gray level;
[0018] Based on the difference between each gray level on each image in each image group and the gray level corresponding to the maximum value of the quantitative characteristic coefficient on that image, as well as the quantitative characteristic coefficient of that gray level, the degree of abnormality in each image group of any MRI sequence is determined.
[0019] Preferably, the step of obtaining the irrelevance index of each pair of target groups under the same MRI sequence based on the difference between abnormalities in each pair of target groups in the same MRI sequence, and the difference between each pair of target groups in each type of MRI sequence and other types of MRI sequences, specifically includes:
[0020] Based on the difference in the degree of abnormality between two target groups under the same MRI sequence, and combined with the difference in the quantitative characteristic coefficients of the two target groups at the same gray level in the MRI sequence, the difference factor of each pair of target groups under each MRI sequence is obtained.
[0021] Based on the difference distribution between the difference factors of each pair of target groups under each MRI sequence and the difference factors of other types of MRI sequences, the irrelevance index of each pair of target groups under each MRI sequence is obtained.
[0022] Preferably, the step of obtaining the difference factor for each pair of target groups under each MRI sequence based on the difference in abnormality degree between each pair of target groups under the same MRI sequence, combined with the difference in the quantitative characteristic coefficients of the two target groups at the same gray level in the MRI sequence, specifically includes:
[0023] For any MRI sequence, the first difference coefficient between each pair of target groups is obtained by acquiring the difference in the degree of abnormality between each pair of target groups under that MRI sequence.
[0024] Based on the difference between the quantitative characteristic coefficients of each pair of target groups at the same gray level in the MRI sequence, a second difference coefficient between each pair of target groups is determined.
[0025] The product of the first and second difference coefficients was normalized to obtain the difference factor for each pair of target groups under this MRI sequence.
[0026] Preferably, the step of obtaining the irrelevance index of each pair of target groups under each MRI sequence based on the difference distribution between the difference factors of each pair of target groups under each MRI sequence and the difference factors of other types of MRI sequences specifically includes:
[0027] For any two target groups, the mean of the difference factors of the two target groups under all types of MRI sequences is obtained as the equilibrium factor; the absolute value of the difference between the difference factors of the two target groups under the target MRI sequence and the equilibrium factor is used as the first deviation coefficient.
[0028] The difference between the difference factor of the two target groups under the reference MRI sequence and the difference factor under the target MRI sequence is used as the second deviation coefficient, where the target MRI sequence is any type of MRI sequence and the reference MRI sequence is any type of MRI sequence other than the target MRI sequence.
[0029] Based on the first and second deviation coefficients, the irrelevance index of the two target groups under the target MRI sequence is determined.
[0030] Preferably, the step of eliminating highly overlapping image groups based on the similarity index between each target group and other target groups to obtain an effective sample set specifically includes:
[0031] For the similarity index between the first target group and each second target group, the proportion of the number of similarity indices greater than the preset similarity threshold is taken as the degree of data overlap of the first target group;
[0032] Among them, the first target group is any target group, and the second target group is any target group other than the first target group;
[0033] Based on the degree of data overlap in each target group, the image data contained in each target group are filtered to obtain an effective sample set.
[0034] Preferably, the step of filtering the image data contained in each target group based on the degree of data overlap to obtain an effective sample set specifically includes:
[0035] The target groups with a data overlap greater than or equal to a preset overlap threshold are removed, while the target groups with a data overlap less than the preset overlap threshold are retained. All types of MRI sequences of the retained target groups constitute the effective sample set.
[0036] Preferably, the step of obtaining the similarity index between each pair of target groups based on the irrelevance index and the difference index corresponding to each pair of target groups specifically includes:
[0037] For any two target groups, the negative correlation coefficient obtained by weighting and summing the differences between the MRI sequences using the negative correlation coefficients of the irrelevance index corresponding to each MRI sequence is the similarity index between the two target groups.
[0038] The embodiments of the present invention have at least the following beneficial effects:
[0039] This invention first collects multiple MRI sequences as initial data for feature analysis, providing a data foundation for subsequent analysis to eliminate redundant data. Then, by analyzing the grayscale ratio of images in each MRI sequence, images with abnormal features are initially screened to obtain target groups, thus eliminating some redundant images in normal states to a certain extent. Further, by analyzing the degree of difference in abnormalities among different target groups under the same type of MRI sequences, and comparing the degree of difference among different types of MRI sequences, an irrelevance index is obtained. This aims to assess whether the difference trend between two groups of images deviates from the overall difference trend, and the reference value of the corresponding image data for similarity feature analysis. Finally, by combining the feature analysis results of the irrelevance index and the degree of difference, the similarity between two target groups is dynamically evaluated, providing a sample similarity that truly reflects the differences in lesion states and avoiding misjudgments of similarity caused by redundant sequences. This invention, by dynamically screening MRI sequences and eliminating redundant sequences that lead to high similarity between samples, constructs a diverse sample set to enhance the model's generalization ability and improve the robustness of cerebral infarction lesion segmentation. Attached Figure Description
[0040] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is a flowchart of the steps of an intelligent segmentation method for cerebral infarction MRI images based on deep learning provided by the present invention;
[0042] Figure 2 This is a flowchart of the steps for obtaining the target group of each MRI sequence provided by the present invention;
[0043] Figure 3 This is a flowchart of the steps for obtaining the irrelevance index provided by the present invention. Detailed Implementation
[0044] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a deep learning-based intelligent segmentation method for cerebral infarction MRI images proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0046] The main objective of this invention is to construct a diverse sample set by dynamically screening MRI sequences and eliminating redundant sequences that lead to high similarity between samples. This helps to avoid overfitting when training neural networks, enhances the model's generalization ability, and improves the robustness of cerebral infarction lesion segmentation results.
[0047] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent segmentation method for cerebral infarction MRI images based on deep learning provided by this invention.
[0048] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based intelligent segmentation method for cerebral infarction MRI images, according to an embodiment of the present invention. The method includes the following steps:
[0049] Step S100: Obtain an image set of each MRI sequence for cerebral infarction symptoms, wherein the image set contains several image groups.
[0050] First, image sets for each MRI sequence are acquired. During the MRI imaging process, by changing the influencing factors of magnetic resonance, multiple different types of MRI sequences can be obtained. These MRI sequence categories include T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and diffusion-weighted imaging (DWI), among others.
[0051] An MRI sequence image set contains several image groups, where each image group refers to the set of images obtained from the same patient's scans. For example, patient A corresponds to one image group in the T1WI sequence image set. Mapping the images within each image group of the same MRI sequence means that the number of images in each different image group is the same. It should be understood that, under different types of sequences and image dimensions, the same patient may correspond to multiple different image groups.
[0052] It should be noted that the scanning acquisition method of MRI sequence images is a well-known technology and will not be described in detail here. In this embodiment, the gray value of each pixel in each image of each MRI sequence has been standardized, so that subsequent analysis can be performed on the gray value features of different types and corresponding dimensions.
[0053] Step S200: Based on the distribution of grayscale values of each image in the image set of each MRI sequence, the image groups with abnormalities are screened to obtain the target group for each MRI sequence.
[0054] Clinical MRI sequences inevitably contain normal samples without lesions, or old samples with completely absorbed lesions and weak signals. If these invalid samples are included directly in subsequent analysis without screening, the high similarity between lesion-free samples will interfere with the assessment of lesion sample diversity when calculating sample similarity. Furthermore, invalid samples will dilute the effective lesion features during subsequent model training, reducing the model's accuracy in identifying real lesions. Screening the target group can eliminate this noise in advance, focusing on samples with lesion potential, while also ensuring the diversity of samples selected subsequently.
[0055] In this regard, such as Figure 2 As shown, the screening steps for the target group of each MRI sequence can be implemented by steps S201 to S204.
[0056] Step S201: Based on the grayscale value of each pixel in each image of each MRI sequence, cluster the pixels of all images in the same MRI sequence to obtain pixel clusters for each MRI sequence.
[0057] In this embodiment, the AP clustering algorithm is used to cluster the gray values of all pixels in all images of the same MRI sequence, and pixels with the same or similar gray values are grouped into the same pixel cluster.
[0058] Step S202: Based on the distribution of gray values in the pixel clusters under each MRI sequence, obtain the gray level of each pixel in each image in the image set of each MRI sequence.
[0059] Specifically, for any MRI sequence, all pixel clusters are arranged in ascending order of gray value, and the gray level of each pixel in each image of the MRI sequence is determined by the arrangement number of the pixel cluster to which each pixel belongs.
[0060] More specifically, taking DWI sequences as an example, the mean of all different gray values within the same pixel cluster is calculated. All pixel clusters in the DWI sequence are then arranged in ascending order of the mean gray value. Each pixel cluster corresponds to a sorting number, and the sorting number is used as the gray level of the pixel in the corresponding pixel cluster.
[0061] The grayscale value range and distribution patterns of different MRI sequences vary greatly (e.g., the absolute grayscale value of a high signal on DWI may differ by several times from that on T2WI). Directly using the original grayscale values to measure sample similarity will result in serious bias. By mapping grayscale values to grayscale levels through AP clustering, the differences in grayscale scale between different sequences can be eliminated, making pixel features comparable across samples and sequences.
[0062] By resetting the corresponding grayscale level for each pixel in each image of each MRI sequence, interference from the original grayscale values can be eliminated, establishing a unified standard for measuring grayscale features. The grayscale level represents the relative level of corresponding grayscale values within the same MRI sequence. This mapping operation eliminates the absolute scale differences of the original grayscale values, transforming the pixel features of different sequences into horizontally comparable level concepts, thus achieving the goal of measuring grayscale features under the same standard.
[0063] Step S203: Based on the proportion of pixels at different gray levels in each image group of each MRI sequence, the degree of abnormality in each image group of each MRI sequence is obtained.
[0064] Since MRI images show the degree of freedom and content of water molecules in the body, when a cerebral infarction occurs, the blood-brain barrier is damaged, and vasogenic edema will appear in the brain. Taking DWI sequence as an example, the edematous area will show a high signal on DWI (the diffusion of water molecules is blocked), so there will be a local bright area (edematous area) in the DWI image.
[0065] In MRI images of cerebral infarction, normal brain tissue accounts for the majority and the components of normal tissue are similar. Based on this characteristic, the pixel with the largest grayscale ratio in each image can be approximated as the normal part. The grayscale difference is used to measure the abnormality of the image, and to a certain extent, some redundant images of normal state are eliminated.
[0066] Specifically, the first step is to obtain the percentage of pixels at each gray level in each image for any MRI sequence as the quantitative feature coefficient of each gray level.
[0067] It should be understood that each different gray level in each image corresponds to a quantitative feature coefficient. For example, the ratio of the number of pixels with gray level H to the total number of pixels in the image is the quantitative feature coefficient of gray level H, which reflects the proportion of pixels with gray level H.
[0068] The second step is to determine the degree of abnormality in each image group of any MRI sequence based on the difference between each gray level on each image in each image group and the gray level corresponding to the maximum value of the quantitative characteristic coefficient on that image, as well as the quantitative characteristic coefficient of that gray level.
[0069] As a concrete example, taking the i-th image group in any MRI sequence as an example, the method for obtaining the abnormality level of the i-th image group can be expressed as follows:
[0070] ;
[0071] in, This indicates the degree of abnormality in the i-th image group of an MRI sequence. This represents the number of images contained in the i-th image group of an MRI sequence. This represents the number of grayscale levels contained in the nth image within the i-th image group of an MRI sequence. This represents the k-th gray level of the n-th image in the i-th image group of an MRI sequence. This represents the gray level corresponding to the maximum value of the quantitative feature coefficient in the nth image of the i-th image group in an MRI sequence. The quantitative feature coefficient represents the k-th gray level of the n-th image in the i-th image group of an MRI sequence. It is a linear normalization function.
[0072] This represents the grayscale features of the normal tissue portion in the nth image, and then... This indicates the degree of deviation and difference between each gray level in the image and the normal baseline. The greater the deviation, the more significant the difference between the gray level's gray-scale characteristics and normal tissue, and the more likely it is to be a lesion area. For example, the high signal response of a lesion in a DWI sequence deviates greatly from the normal baseline.
[0073] The greater the deviation of the grayscale features from normal, and the larger the proportion of that grayscale level in the image (such as a large lesion area), the higher the degree of abnormality. The degree of abnormality of each image group reflects the degree of lesion abnormality in the images within that group. The larger the value, the greater the likelihood that the images in that group contain lesions, and the greater the contribution to the sample diversity of subsequent lesion segmentation.
[0074] Step S204: The image group corresponding to the abnormality level being greater than or equal to the preset abnormality threshold is selected as the target group of the corresponding MRI sequence.
[0075] In this embodiment, the anomaly threshold is set to 0.3. When the anomaly level of an image group in each MRI sequence is greater than or equal to the anomaly threshold, it indicates that the image group is more likely to have lesion features, providing a data foundation for constructing diverse samples. When the anomaly level of an image group in each MRI sequence is less than the anomaly threshold, it indicates that the images in the image group are less likely to have lesion features. In this case, such image data is removed, and no subsequent similarity feature analysis is performed, which reduces the redundancy of image data to a certain extent.
[0076] By screening the image groups of each MRI sequence according to the degree of abnormality, the initial screening of the samples was achieved, and the target groups in the samples that may contain lesions were extracted. In subsequent steps, the image sequences of each target group were analyzed.
[0077] Step S300: Based on the degree of difference between abnormalities in the same MRI sequence for each two target groups, and the degree of difference between each two target groups in each MRI sequence and other types of MRI sequences, obtain the irrelevance index of each two target groups in the same MRI sequence.
[0078] MRI images exhibit various sequences, each reflecting different physiochemical states of brain tissue, such as content and degrees of freedom. The appearance of cerebral infarction lesions also varies at different stages. For example, in the hyperacute phase of cerebral infarction, cytotoxic edema occurs, resulting in significantly elevated edema areas on DWI sequences. In the chronic phase, cerebral softening and cystic degeneration occur instead of edema, and the overall appearance of DWI sequences is more similar to normal cells. The manifestations of DWI sequences vary considerably across different stages.
[0079] Therefore, different MRI sequences exhibit significant differences in their ability to distinguish different stages of cerebral infarction lesions. Taking DWI sequences as an example, the differences in image features with lesion stage are obvious: if two sets of images show similar DWI sequence characteristics, it can be determined that the lesion stages corresponding to these two sets of images are similar, meaning that this sequence has a good effect on distinguishing lesion stages. Conversely, some MRI sequences show relatively small differences in image features across different stages of cerebral infarction lesions. For these types of sequences, even if two sets of images show similar characteristics on this sequence, it cannot be used as a basis for determining that the lesion stages corresponding to the two sets of images are similar.
[0080] Based on the above characteristics, it is necessary to analyze the reference value of different sequences as criteria for measuring sample similarity by combining the relative differences between the images in each sequence. Specifically, when two sets of images show significant differences in most sequences but only show high similarity in one sequence, the sequence is likely a sequence that shows similarity in different lesion stages of cerebral infarction, and its reliability as a criterion for measuring sample similarity is correspondingly low.
[0081] To address this, the irrelevance index is used to assess whether the trend of difference between the two sets of images deviates from the overall trend of difference. The greater the deviance, the lower the reference value of the corresponding data. Figure 3 As shown, the irrelevance index can be obtained by steps S301 and S302.
[0082] Step S301: Based on the difference in the degree of abnormality between two target groups under the same MRI sequence, and combined with the difference in the quantitative characteristic coefficients of the two target groups at the same gray level in the MRI sequence, the difference factor of each pair of target groups under each MRI sequence is obtained.
[0083] Specifically, the analysis process for the difference factors between every two target groups is the same for each MRI sequence. This embodiment uses any MRI sequence as an example, such as the DWI sequence.
[0084] The first step is to obtain the first difference coefficient between each pair of target groups by acquiring the difference in the degree of abnormality between each pair of target groups under that MRI sequence for any given MRI sequence.
[0085] As a specific example, the absolute value of the difference between the abnormality levels of target group A and target group B in the DWI sequence is used to obtain the first difference coefficient between target group A and target group B in the DWI sequence.
[0086] The first difference coefficient reflects the difference in the degree of abnormality of lesion features exhibited by each two target groups in the same DRI sequence.
[0087] The second step is to determine the second difference coefficient between each pair of target groups based on the difference between the quantitative characteristic coefficients at the same gray level in the MRI sequence.
[0088] As a concrete example, calculate the absolute value of the difference between the quantitative feature coefficients of target group A and target group B at the k-th gray level in the m-th image of the DWI sequence. The absolute value of the difference between target group A and target group B at all the same gray levels in the DWI sequence. The differences were accumulated to obtain the second difference coefficient between target group A and target group B under DWI sequences. ,in This represents the maximum number of gray levels contained in the m-th image corresponding to target group A and target group B in the DWI sequence. This indicates the number of images contained in target group A and target group B in the DWI sequence.
[0089] The grayscale value of a pixel represents the physicochemical state of water molecules within the tissue. A greater difference in the size and proportion of grayscale distribution indicates a greater difference in the stage of lesion development and the lesion area. The second difference coefficient reflects the difference in the size of lesion features between two target groups in the same MRI sequence.
[0090] The third step is to normalize the product between the first and second difference coefficients to obtain the difference factor for each pair of target groups under this MRI sequence.
[0091] As a concrete example, the minimax normalization method can be used to normalize the product of the first and second difference coefficients between every two target groups in the DWI sequence to obtain the difference factor between every two target groups in the DWI sequence.
[0092] The differential factor, combining the results of differential analysis of lesion features from two aspects, measures the difference in lesion characteristics between two target groups in the same MRI sequence. The differential factor reflects the difference in lesion features between the two target groups in that sequence; a larger value indicates a more significant difference. Simultaneously, the differential factor characterizes the degree of difference between abnormalities in two target groups within the same MRI sequence.
[0093] Step S302: Based on the difference distribution between the difference factors of each pair of target groups under each MRI sequence and the difference factors of other types of MRI sequences, obtain the irrelevance index of each pair of target groups under each MRI sequence.
[0094] For any two target groups, if the difference is small under DWI sequence but large under other sequences, it indicates that the difference correlation under DWI sequence is low. This suggests that the tissue water status shown in the images of this sequence may be similar at various stages of the development of cerebral infarction lesions. Ultimately, it indicates that the similarity between the two target groups under this sequence is not a true similarity of lesion features. Such similarity features have low reference value for the similarity calculation of the two target groups.
[0095] Based on this, the irrelevance index of each pair of target groups under each MRI sequence was quantified to evaluate the reference value of the corresponding image data for the process of measuring the similarity between the two groups.
[0096] Specifically, for any two target groups, the mean of the difference factors of the two target groups under all types of MRI sequences is obtained as the equilibrium factor; the absolute value of the difference between the difference factor of the two target groups under the target MRI sequence and the equilibrium factor is used as the first deviation coefficient; the difference between the difference factor of the two target groups under the reference MRI sequence and the difference factor under the target MRI sequence is used as the second deviation coefficient, where the target MRI sequence is any type of MRI sequence and the reference MRI sequence is any type of MRI sequence other than the target MRI sequence; based on the first deviation coefficient and the second deviation coefficient, the irrelevance index of the two target groups under the target MRI sequence is determined.
[0097] As a concrete example, if the r-th MRI sequence is taken as the target MRI sequence, then all other MRI sequences except the r-th MRI sequence are reference MRI sequences. For target group A and target group B, the method for obtaining the irrelevance index under the target MRI sequence can be expressed by the formula:
[0098] ;
[0099] in, denoted as the irrelevance index between target group A and target group B under the target MRI sequence, where r represents the r-th MRI sequence; This represents the difference factor between target group A and target group B under the x-th reference MRI sequence. This represents the difference factor between target group A and target group B under the target MRI sequence (the r-th MRI sequence). This represents the mean of the difference factors between target group A and target group B across all MRI sequences, also known as the equilibrium factor. This represents the total number of reference MRI sequences corresponding to the target MRI sequence, which is the total number of all types of MRI sequences minus 1. Adding 1 to the denominator prevents the difference between the difference factor and the equilibrium factor from being zero, thus ensuring a denominator of zero and affecting the calculation results.
[0100] The first deviation factor represents the deviation between the feature differences of the two target groups under the MRI sequence and the overall feature differences. The larger the value, the greater the degree of deviation of the feature differences in other sequences from the overall difference state. The first deviation factor essentially characterizes the confidence weight corresponding to the feature differences in other sequences.
[0101] The second deviation factor represents the difference in features between the two target groups under the reference MRI sequence and the difference in features under the target MRI sequence. The larger the value, the greater the feature difference in other sequences and the smaller the feature difference in the current sequence. In this case, the difference between the two is greater.
[0102] For target group A and target group B, the comparison results of all types of sequences are considered to avoid the influence of abnormal performance of a single type of sequence on the judgment. If most other types of MRI sequences show that the difference of the target MRI sequence is much smaller than that of other types of MRI sequences, the final irrelevance index value will be significantly increased. In this case, the target MRI sequence is an irrelevant sequence with low reference value for judging the similarity between the two target groups.
[0103] The irrelevance index characterizes the reference value of image data of two target groups in a target MRI sequence for the similarity between the two target groups. Essentially, it dynamically identifies disjointed sequences by comparing the differences between individual sequences and the overall sequence difference trend. When the difference performance of a certain sequence does not match the difference level of most sequences and the overall difference level (e.g., most sequences have large differences, but only this sequence has small differences), the larger the irrelevance index value of the sequence, the smaller the reference value, and the more likely the sequence is to be an irrelevant sequence. Its weight will be reduced when calculating the similarity of samples in the future to avoid misjudgment of similarity caused by this sequence, thus providing a basis for constructing a diverse sample set.
[0104] Step S400: Obtain the similarity index between each pair of target groups based on the irrelevance index and the difference index corresponding to each pair of target groups; based on the similarity index between each target group and other target groups, remove highly overlapping image groups to obtain an effective sample set, which is used to train the intelligent segmentation neural network.
[0105] First, the irrelevance index characterizes the degree of reference value of a sample in similarity calculation. By using the irrelevance index to weight the differences between samples to calculate similarity, it is possible to dynamically adjust the contribution weight of image data of two target groups under each MRI sequence to the sample similarity judgment, reduce the influence of MRI sequence data with low reference value, amplify the influence of key sequences (MRI sequence data with high reference value), and finally obtain sample similarity that can truly reflect the differences in lesion status, avoiding misjudgment of similarity caused by redundant sequences.
[0106] Specifically, for any two target groups, the negative correlation coefficient obtained by weighted summing of the differences corresponding to each MRI sequence using the irrelevance index for each MRI sequence is the similarity index between the two target groups. It should be noted that the two target groups correspond to two different patients, and the degree of similarity between the two patients is assessed by analyzing the data reference value of these two patients across various sequences.
[0107] As a concrete example, taking target group A and target group B as examples, the method for obtaining the similarity index can be expressed by the formula:
[0108] ;
[0109] in, This represents a similarity index between target group A (corresponding patients) and target group B (corresponding patients). This represents the similarity reference weights of target group A and target group B corresponding to the r-th MRI sequence. , This represents the degree of irrelevance between target group A and target group B under the r-th MRI sequence. This represents the difference factor between target group A and target group B under the r-th MRI sequence. This indicates the number of different types of MRI sequences. This represents an exponential function with the natural constant e as its base.
[0110] The similarity metric is measured by weighting sequences with lower irrelevance, resulting in a corrected similarity score. This avoids inflated overall similarity scores due to sequences with less influence on the overall similarity. The similarity metric reflects a more realistic picture similarity between two target groups.
[0111] Then, the similarity features between different target groups are analyzed. When a target group has high similarity with many other target groups, it is necessary to extract samples of the highly similar target group from the training samples to avoid overfitting of the model.
[0112] The first step is to obtain the proportion of the number of similarity indicators greater than a preset similarity threshold between the first target group and each second target group as the data overlap degree of the first target group; where the first target group is any target group and the second target group is any other target group besides the first target group.
[0113] As a concrete example, if there are M target groups, then M-1 similarity indicators are calculated between the first target group and each second target group. These similarity indicators are compared with a similarity threshold, and the ratio of the number of similarity indicators greater than the similarity threshold to M-1 is calculated to obtain the data overlap degree of the first target group. In this embodiment, the similarity threshold is set to 0.8, but the implementer can set it according to the specific implementation scenario.
[0114] The data overlap rate of the first target group indicates the proportion of images in the first target group that are highly similar to images in other target groups. A higher overlap rate indicates greater overlap, which may lead to overfitting in the model, thus requiring its removal from the training samples. Conversely, a lower overlap rate indicates less overlap, making the first target group a more effective training sample and satisfying the requirement for sample richness; therefore, it does not need to be removed from the training samples.
[0115] The second step is to filter the image data contained in each target group based on the degree of data overlap to obtain a valid sample set.
[0116] Specifically, target groups with a data overlap greater than or equal to a preset overlap threshold are removed, while target groups with a data overlap less than the preset overlap threshold are retained. All types of MRI sequences in the retained target groups constitute the valid sample set. In this embodiment, the overlap threshold is set to 0.2, but the implementer can set it according to the specific implementation scenario.
[0117] When the overlap of data in the target group is less than the overlap threshold, it indicates that the image data of the target group in each MRI sequence is a valid sample. The image data of each MRI sequence contained in all retained target groups constitute a valid sample set. The valid sample set is used to train a neural network for intelligent segmentation of cerebral infarction MRI image lesions, realizing a deep learning-based intelligent segmentation method for cerebral infarction MRI image lesions. The neural network used for image segmentation is a U-Net neural network, and the method for training the network is a well-known technique and is not the focus of this embodiment, so it will not be elaborated upon here.
[0118] In summary, this invention constructs a diverse sample set by dynamically screening MRI sequences and eliminating redundant sequences that lead to high similarity between samples, thereby enhancing the model's generalization ability and improving the robustness of cerebral infarction lesion segmentation.
[0119] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for intelligent segmentation of cerebral infarction MRI images based on deep learning, characterized in that, The method includes the following steps: The image set of each MRI sequence for cerebral infarction is obtained, wherein the image set contains several image groups, and the same image group refers to the set of images obtained from the scan of the same patient; Based on the distribution of grayscale values in each image within the image set of each MRI sequence, image groups with anomalies are screened to obtain the target group for each MRI sequence; Based on the degree of difference between abnormalities in the same MRI sequence for each two target groups, and the degree of difference between each two target groups in each MRI sequence and other types of MRI sequences, the irrelevance index of each two target groups in the same MRI sequence is obtained. The similarity index between each pair of target groups is obtained based on the irrelevance index and the difference index corresponding to each pair of target groups; based on the similarity index between each target group and other target groups, highly overlapping image groups are eliminated to obtain an effective sample set, which is used to train the intelligent segmentation neural network. The method for obtaining the target group for each MRI sequence specifically includes: Based on the grayscale value of each pixel in each image of each MRI sequence, the pixels of all images in the same MRI sequence are clustered to obtain pixel clusters for each MRI sequence. Based on the distribution of gray values in pixel clusters under each MRI sequence, the gray level of each pixel in each image in the image set of each MRI sequence is obtained; The degree of abnormality in each image group of each MRI sequence is obtained by considering the proportion of pixels at different gray levels in each image group of each MRI sequence. The image group corresponding to the abnormality level being greater than or equal to the preset abnormality threshold is designated as the target group for the corresponding MRI sequence; The method for obtaining the irrelevance index of each pair of target groups under the same MRI sequence specifically includes: Based on the difference in the degree of abnormality between two target groups under the same MRI sequence, and combined with the difference in the quantitative characteristic coefficients of the two target groups at the same gray level in the MRI sequence, the difference factor of each pair of target groups under each MRI sequence is obtained. Based on the difference distribution between the differential factors of each pair of target groups under each MRI sequence and the differential factors of other types of MRI sequences, the irrelevance index of each pair of target groups under each MRI sequence is obtained, including: For any two target groups, the mean of the difference factors of the two target groups under all types of MRI sequences is obtained as the equilibrium factor; the absolute value of the difference between the difference factors of the two target groups under the target MRI sequence and the equilibrium factor is used as the first deviation coefficient. The difference between the difference factor of the two target groups under the reference MRI sequence and the difference factor under the target MRI sequence is used as the second deviation coefficient, where the target MRI sequence is any type of MRI sequence and the reference MRI sequence is any type of MRI sequence other than the target MRI sequence. Based on the first and second deviation coefficients, the irrelevance index of the two target groups under the target MRI sequence is determined.
2. The method for intelligent segmentation of cerebral infarction MRI images based on deep learning according to claim 1, characterized in that, The step of obtaining the gray level of each pixel in each image of each MRI sequence image set based on the gray value distribution in pixel clusters under each MRI sequence specifically includes: For any MRI sequence, all pixel clusters are arranged in ascending order of gray value. The gray level of each pixel in each image of the MRI sequence is determined by the arrangement number of the pixel cluster to which each pixel belongs.
3. The method for intelligent segmentation of cerebral infarction MRI images based on deep learning according to claim 1, characterized in that, The degree of abnormality in each image group of each MRI sequence is determined based on the proportion of pixels at different gray levels in each image group within each MRI sequence. Specifically, this includes: For any MRI sequence, the percentage of pixels at each gray level in each image is obtained as the quantitative feature coefficient of each gray level; Based on the difference between each gray level on each image in each image group and the gray level corresponding to the maximum value of the quantitative characteristic coefficient on that image, as well as the quantitative characteristic coefficient of that gray level, the degree of abnormality in each image group of any MRI sequence is determined.
4. The intelligent segmentation method for cerebral infarction MRI images based on deep learning according to claim 1, characterized in that, The method involves obtaining the difference factor for each pair of target groups under each MRI sequence based on the difference in abnormality between two target groups under the same MRI sequence, combined with the difference in the quantitative characteristic coefficients of the two target groups at the same gray level in the MRI sequence. Specifically, this includes: For any MRI sequence, the first difference coefficient between each pair of target groups is obtained by acquiring the difference in the degree of abnormality between each pair of target groups under that MRI sequence. Based on the difference between the quantitative characteristic coefficients of each pair of target groups at the same gray level in the MRI sequence, a second difference coefficient between each pair of target groups is determined. The product of the first and second difference coefficients was normalized to obtain the difference factor for each pair of target groups under this MRI sequence.
5. The method for intelligent segmentation of cerebral infarction MRI images based on deep learning according to claim 1, characterized in that, The step of eliminating highly overlapping image groups based on the similarity index between each target group and other target groups to obtain an effective sample set specifically includes: For the similarity index between the first target group and each second target group, the proportion of the number of similarity indices greater than the preset similarity threshold is taken as the degree of data overlap of the first target group; Among them, the first target group is any target group, and the second target group is any target group other than the first target group; Based on the degree of data overlap in each target group, the image data contained in each target group are filtered to obtain an effective sample set.
6. The method for intelligent segmentation of cerebral infarction MRI images based on deep learning according to claim 5, characterized in that, The step of filtering the image data contained in each target group based on the degree of data overlap to obtain an effective sample set specifically includes: The target groups with a data overlap greater than or equal to a preset overlap threshold are removed, while the target groups with a data overlap less than the preset overlap threshold are retained. All types of MRI sequences of the retained target groups constitute the effective sample set.
7. The intelligent segmentation method for cerebral infarction MRI images based on deep learning according to claim 1, characterized in that, The similarity index between each pair of target groups is obtained based on the irrelevance index and the difference index corresponding to each pair of target groups, specifically including: For any two target groups, the negative correlation coefficient obtained by weighting and summing the differences between the MRI sequences using the negative correlation coefficients of the irrelevance index corresponding to each MRI sequence is the similarity index between the two target groups.