CN114359637A - Brain medical image classification method and device - Google Patents
Brain medical image classification method and device Download PDFInfo
- Publication number
- CN114359637A CN114359637A CN202210021393.9A CN202210021393A CN114359637A CN 114359637 A CN114359637 A CN 114359637A CN 202210021393 A CN202210021393 A CN 202210021393A CN 114359637 A CN114359637 A CN 114359637A
- Authority
- CN
- China
- Prior art keywords
- medical image
- brain medical
- brain
- module
- sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention relates to a brain medical image classification method and device. The brain medical image classification method comprises the following steps: acquiring a brain medical image sequence to be identified; preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data; inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module. According to the brain medical image classification method, the image classification model introduces an attention mechanism, so that the expression of effective information is improved, the spatial information in the sequence is accurately utilized, the characteristic extraction effect is improved, and the model can distinguish whether the given brain medical image (sequence) is consistent with the stroke medical image (sequence) or not with high accuracy.
Description
Technical Field
The invention relates to the field of computer vision, in particular to a brain medical image classification method and device.
Background
At present, the research on intelligent diagnosis of stroke at home and abroad is mainly divided into two aspects of stroke risk prediction and stroke focus segmentation. (1) In the aspect of stroke screening, a stroke screening model is established by using three algorithms of Artificial Neural Network (ANN), logistic regression and decision tree for statistical data of family history, physical quality, motion condition and the like of residents in 2016 by using Huangxianxian, Yangyijie and the like; (2) in the aspect of stroke focus segmentation, Yu and the like use a U-Net neural network to predict the infarction position of an acute stroke patient. The experimental result shows that after multiple times of verification, the median overlapping area of the model prediction infarction range and the actual infarction range reaches 92%. The yellow pears and lulos use a depth residual algorithm to improve the U-Net neural network and insert long-distance dependencies in non-local block coding profiles. And (3) segmenting the ischemic stroke focus by using the improved U-Net network. The DICE coefficient of the residual U-Net obtained as a result was 0.29. + -. 0.23 (the DICE coefficient ranged from 0 to 1, and closer to 1 indicates better segmentation result). Yangho adds a plurality of jump connections and introduces a hole convolution algorithm on the basis of MSDF-Net, and the average DICE coefficient of a result model in a chronic stroke lesion segmentation task reaches 0.58. (3) In the aspect of brain medical image classification, Gao.X et al use a convolutional neural network to distinguish three types of brain medical images, namely Alzheimer's disease, tumor and normal aging, wherein the 2D CNN reaches 86.32% accuracy rate, and the 3D CNN reaches 85.26% accuracy rate. Prevedello.L.M and other people use an artificial intelligence algorithm to judge whether the brain medical image contains stroke, and research results show that the sensitivity and the specificity of the artificial intelligence algorithm are 62% and 95%. Titano.j.j et al used an EMR prioritization algorithm and a 3D CNN network algorithm to determine whether medical images of the brain contain strokes and ponding. The accuracy of the EMR prioritization algorithm is 55%, the sensitivity is 16%, and the specificity is 73%; the accuracy of the 3D CNN network algorithm was 71%, sensitivity 18%, specificity 95%.
As can be seen from the analysis of the current state of the art, most of the studies on assisted diagnosis of stroke by deep learning are in the field of image segmentation by deep learning, but the studies on the field of image classification by deep learning are less. However, the segmentation network may wrongly divide the normal region into the focus region, so that the patient is often misdiagnosed as a stroke patient only by relying on the segmentation network to assist clinical diagnosis, and the study on image classification of medical images in stroke is also very important, and the classification network and the segmentation network jointly applied to the auxiliary diagnosis can achieve a better effect. In addition, a typical neural network such as U-Net uses a 2D convolution kernel for sampling, and spatial information of a stroke focus in a medical image sequence cannot be well utilized, which affects a segmentation effect. In a few studies of the existing deep learning-based stroke medical image classification, 3D convolutional network structures basically used by students are mostly simple stacks of convolutional layers, batch normalization layers and maximum pooling layers, the classification effect is general, and the model structure can still be continuously optimized.
Disclosure of Invention
Based on this, the present invention provides a method and an apparatus for classifying brain medical images, in which an image classification model is used to introduce an attention mechanism (hereinafter abbreviated as SE-3D ResNet), so as to improve the expression of effective information, and accurately utilize spatial information of a stroke lesion in a sequence to improve a feature extraction effect, so that the model can distinguish whether a given brain medical image (sequence) is consistent with a stroke medical image (sequence) with a high accuracy.
In a first aspect, the present invention provides a method for classifying brain medical images, the method comprising the following steps:
acquiring a brain medical image sequence to be identified;
preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data;
inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
Further, inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence, including:
inputting the brain medical image sequence with the normalized data into the input module for convolution processing to obtain a primary effective characteristic diagram;
inputting the preliminary effective feature map into the residual error learning module for processing to obtain a key feature map;
and inputting the key feature map into the output module for classification processing to obtain a classification result of the brain medical image sequence.
Further, the residual learning module comprises 4 residual modules incorporating attention mechanism;
each residual module comprises a first convolution layer, a first BN layer, a first ReLU active layer, a second convolution layer, a second BN layer, an SE module and a second ReLU active layer which are connected in sequence;
the output of the SE module is added with the original information of the input module through jump connection, and then is used as the output of the SE-Res module through a ReLU activation function.
Further, inputting the preliminary effective feature map into the residual error learning module for processing to obtain a key feature map, including:
processing the preliminary effective characteristic diagram x input into the residual error module to obtain a key characteristic diagram U by using the following formulac：
Uc＝SE(δ(Conv3×3×3(L(B(Conv3×3×3(x)))))+x
SE(in)＝σ(W2·δ(W1·Fsq(in))
Where x represents the feature map of the most input residual module, B () represents the Batch Normalization operation, δ () represents the Leaky ReLU activation function, in represents the feature map of the input SE module, FsqThe function represents a global average pooling operation; w1 and W2 represent parameters of two fully connected layers, respectively; σ (.) denotes Sigmoid activation function.
Further, preprocessing the brain medical image sequence to obtain a data-normalized brain medical image sequence, including:
reducing the original data of 512 × 512 pixels into a gray image of 224 × 224 pixels by using a nearest neighbor interpolation method;
processing salt-pepper noise and speckle noise in the brain medical image by using median filtering and denoising, so that the low-density shadow of the target region becomes a more continuous low-density shadow surface;
and improving the contrast of the brain medical image by using histogram equalization to obtain a brain medical image sequence with data normalization.
Further, improving the contrast of the brain medical image using histogram equalization includes:
calculating the kth gray level r in a brain medical image by using the following formulakProbability of occurrence:
where n represents the total number of pixels in the image, L represents the number of gray levels in the image, nkRepresenting a grey level of rkThe number of pixels of (a); r iskIs the k-th gray level, p, in the brain medical imagek(rk) Representing the kth gray level r in the medical image of the brainkThe probability of occurrence.
Calculating the transformed gray level s of each pixel using the following histogram equalized discrete gray level transformation function formulak：
Wherein s iskThe gray level of each pixel after equalization is between 0 and 1, rjIs the j-th gray level, pr(rj) Is rjProbability of occurrence of gray scale, njTo a gray level of rjThe number of pixels of (a);
the gray level in the original image is rkEach pixel of (a) is mapped to a new image after histogram equalization as skTo obtain the brain medical image sequence with data normalization.
Further, the training step of the image classification model comprises:
acquiring a brain medical image CT sequence of a cerebral apoplexy patient;
preprocessing the brain medical image CT sequence of the cerebral apoplexy patient to obtain a data set containing the brain medical image CT sequence of the cerebral apoplexy patient with data normalization;
dividing the data set into a training set and a validation set;
and training the image classification model by using the training set, and selecting parameters of the entity recognition model by using the verification set to obtain the trained image classification model.
In a second aspect, the present invention also provides a brain medical image classification device, including:
the image acquisition module is used for acquiring a brain medical image sequence to be identified;
the preprocessing module is used for preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data;
the classification module is used for inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
The brain medical image classification method and device provided by the invention creatively and combinatively use two algorithms of a median filtering denoising algorithm and a histogram equalization algorithm as experimental data preprocessing algorithms, provide a 3D convolutional neural network based on a residual network framework, and introduce an attention mechanism (hereinafter abbreviated as SE-3D ResNet), wherein the algorithm designs a coping scheme of gradient disappearance based on a traditional convolutional neural network algorithm, integrates the attention mechanism to inhibit invalid information, improves the expression of effective information, accurately utilizes the spatial information of a stroke focus in a sequence to improve the feature extraction effect, and enables a model to distinguish whether a given brain medical image (sequence) is the stroke medical image (sequence) with higher accuracy. Experimental data show that the classification effect of the model can be improved to a greater extent by combining three methods of using a 3D convolution kernel, introducing an attention mechanism and introducing a residual network framework.
For a better understanding and practice, the invention is described in detail below with reference to the accompanying drawings.
Drawings
Fig. 1 is a schematic flow chart of a method for classifying brain medical images according to the present invention;
FIG. 2 is a schematic diagram illustrating comparison of effects before and after local histogram equalization processing on a medical brain image according to an embodiment of the present invention;
FIG. 3 is a block diagram of an image classification model used in one embodiment of the present invention;
FIG. 4 is a block diagram of a residual module used in one embodiment of the present invention;
fig. 5 is a schematic structural diagram of a brain medical image classification apparatus according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.
It should be understood that the embodiments described are only some embodiments of the present application, and not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without any creative effort belong to the protection scope of the embodiments in the present application.
The terminology used in the embodiments of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the embodiments of the present application. As used in the examples of this application and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.
When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the application, as detailed in the appended claims. In the description of the present application, it is to be understood that the terms "first," "second," "third," and the like are used solely to distinguish one from another and are not necessarily used to describe a particular order or sequence, nor are they to be construed as indicating or implying relative importance. The specific meaning of the above terms in the present application can be understood by those of ordinary skill in the art as appropriate.
Further, in the description of the present application, "a plurality" means two or more unless otherwise specified. "and/or" describes the association relationship of the associated objects, meaning that there may be three relationships, e.g., a and/or B, which may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship.
In view of the problems in the background art, the present application provides a method for classifying brain medical images, as shown in fig. 1, the method including the following steps:
s01: and acquiring a brain medical image sequence to be identified.
In a specific embodiment, the original brain medical image sequence contains 16 to 26 slices, the original images are affected by equipment, environment and other factors, the slices of the front and back parts are blurred, and preferably, the slices of the middle part are reserved, and 10 slices from the 5 th slice to the 14 th slice are classified.
S02: and preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data.
In a preferred embodiment, step S02 includes the following sub-steps:
s021: original data of 512 × 512 pixels is reduced to a grayscale image of 224 × 224 pixels using a nearest neighbor interpolation method.
S022: salt and pepper noise and speckle noise in the brain medical image are processed by using median filtering denoising, so that the low-density shadow of the target region becomes a more continuous low-density shadow surface.
S023: and improving the contrast of the brain medical image by using histogram equalization to obtain a brain medical image sequence with data normalization.
When the global histogram equalization processing image is used, the image may be influenced by the non-target area, so that the contrast of the target area is reduced, and the contrast of the non-target area is improved. The local histogram equalization is to set the size of the processing region, divide the image into a plurality of regions with set sizes, and perform histogram equalization on each region.
Specifically, the histogram equalization step includes:
s0231: calculating the kth gray level r in a brain medical image by using the following formulakProbability of occurrence:
where n represents the total number of pixels in the image, L represents the number of gray levels in the image, nkRepresenting a grey level of rkThe number of pixels of (a); r iskIs the k-th gray level, p, in the brain medical imagek(rk) Representing the kth gray level r in the medical image of the brainkThe probability of occurrence.
S0232: calculating the transformed gray level s of each pixel using the following histogram equalized discrete gray level transformation function formulak：
Wherein s iskThe gray level of each pixel after equalization is between 0 and 1, rjIs the j-th gray level, pr(rj) Is rjProbability of occurrence of gray scale, njTo a gray level of rjThe number of pixels of (a);
s0233: the gray level in the original image is rkEach pixel of (a) is mapped to a new image after histogram equalization as skTo obtain the brain medical image sequence with data normalization.
Fig. 2 shows a comparison graph of the effect before and after local histogram equalization processing on a denoised CT medical image, where fig. 2(a) is an effect graph before histogram equalization processing, and fig. 2(b) is an effect graph after histogram equalization processing
S03: inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
In a specific embodiment, the structure of the SE-3D ResNet model is shown in fig. 3, and includes 1 input module, 4 residual learning modules and output modules connected in sequence. The input module comprises a convolution layer with convolution kernel size of 7 multiplied by 7, a 3D BN layer, a ReLU and a maximum pooling layer. The input feature map is (Batch, Channel)in,Din,Hin,Win) Selection of textThe default value of disparity is 1, and the formula of the size of the feature map after the convolution of the three-dimensional image is shown as the following formula.
The output feature map has a size of (Batch, Channel)out,Dout,Hout,Wout) Where D denotes the depth of the feature map, H denotes the height of the feature map, W denotes the width of the feature map, padding denotes the implicit padding on both sides of the input, the step size of the stride convolution kernel, and kernel _ size denotes the convolution kernel size.
As shown in fig. 4, each residual module includes a first convolution layer, a first BN layer, a first ReLU active layer, a second convolution layer, a second BN layer, an SE module, and a second ReLU active layer, which are connected in sequence. The output of the SE module is added with the original information of the input module through jump connection, and then is used as the output of the SE-Res module through a ReLU activation function. After the brain medical image sequence passes through the input module, the brain medical image sequence enters a Residual error module (SE-Residual Unit) combined with an attention mechanism to obtain the key characteristic map information U of the strokec。
Uc＝SE(δ(Conv3×3×3(L(B(Conv3×3×3(x)))))+x (4)
SE(in)＝σ(W2·δ(W1·Fsq(in))) (5)
Where x represents the feature map of the most input residual module, B () represents the Batch Normalization operation, δ () represents the Leaky ReLU activation function, in the feature map of the input SE module, FsqThe function represents a global average pooling operation; w1And W2Respectively representing two full connectionsParameters of the layer; σ (.) denotes Sigmoid activation function.
The output module comprises a global average pooling layer and a full-connection layer, and key characteristic graph information U of the cerebral apoplexy is obtained through calculation of the residual learning modulecAnd the data enters an output module, and the characteristic dimensionality is reduced through an average pooling layer, so that the parameters and the calculated amount of a full connection layer are reduced. And mapping the learned features to a sample marking space through a full connection layer for classification to obtain a final network output Out, wherein the network output indicates whether the sequence is consistent with a brain medical image sequence of a cerebral apoplexy patient, T indicates consistency, and F indicates inconsistency. The classification flow chart and formula are as follows:
Out＝Fc(Fsq(Uc)) (6)
wherein FsqRepresenting a global average pooling operation, FcRepresenting a fully connected layer.
In a preferred embodiment, the training step of the image classification model includes:
s11: the method comprises the steps of obtaining a brain medical image CT sequence of a cerebral apoplexy patient.
Specifically, the original data used by the invention is a brain medical image CT sequence of 831 patients, each brain medical image sequence of 831 patients contains 16 to 26 slices, after part of problem data is eliminated, 11197 slices are totally included, each slice has the size of 512 × 512 pixels, and is stored in a jpeg format. The taken images are affected by factors such as equipment and environment, the slices in the front and rear parts are fuzzy, so the middle part slices are reserved in the text, and 10 slices from the 5 th slice to the 14 th slice are classified.
S12: and preprocessing the brain medical image CT sequence of the cerebral apoplexy patient to obtain a data set of the brain medical image CT sequence of the cerebral apoplexy patient with data normalization.
As mentioned above, the preprocessing step includes using nearest neighbor interpolation to reduce the original data of the original image with 512 × 512 pixels into a gray image with 224 × 224 pixels, performing median filtering and denoising to obtain a preliminary image, and performing histogram equalization to obtain a final input image.
S13: the data set is divided into a training set and a validation set.
In the training of the 3D residual neural network, 191 patient brain medical image sequences are divided into a verification set, and 640 patient brain medical image sequences are divided into a training set.
S14: and training the image classification model by using the training set, and selecting parameters of the entity recognition model by using the verification set to obtain the trained image classification model.
The preprocessing strategy of the invention can relieve the problem of calculation overhead of the existing medical image classification method, and maintain the texture and edge characteristics of data to a certain extent, so as to obtain the CT image with high contrast of focus and normal human tissue. The results of the classification experiment using the partial neural network algorithm of the preprocessed data set and the original data set are shown in the following tables 1(a) and 1(b), respectively:
TABLE 1(a) Experimental results for raw data set
TABLE 1(b) Experimental results on data-preprocessed data set
The above experimental results show that the classification effect can be effectively improved by the combination of the median filtering denoising algorithm and the histogram equalization algorithm for data preprocessing.
Compared with the existing 3D CNN model, the SE-3D ResNet model used by the invention introduces a residual error module, enhances information interaction among different channels, promotes more effective feature projection convergence, solves the problem of gradient disappearance that the weight of a front layer is difficult to update in the back propagation process to a great extent, and accelerates the speed of network training. According to the CT image knowledge of the cerebral apoplexy, the cerebral apoplexy focus basically exists in the slice in the middle of the sequence, so the front and the last slices are possibly useless information. In order to show the effectiveness and the reasonableness of the proposed module, the image classification model used by the invention is subjected to a transverse contrast experiment with a 2D and 3D convolutional neural network.
In the 2D convolutional neural network experiment, the predicted effect of each typical 2D convolutional neural network and the combination of attention mechanism is shown in table 2. The upper half of table 2 shows that the residual convolutional neural network performs well on the data set, and can better fit the small data set used herein with less parameter quantity, which means that the residual module better solves the over-fitting problem and the gradient disappearance problem by using a jump connection and batch normalization layer, and verifies the correctness of using the residual neural network structure as a basic frame; the lower half of table 2 shows that the attention mechanism can make the model have better classification effect by suppressing the extraneous features.
TABLE 2 Pre-and post-prediction Effect of each original 2D network in combination with attention mechanism
Bold represents the model with the best predicted effect in the table and its predicted effect.
Table 3 lateral contrast experimental results of 3D convolutional neural network herein
Bold represents the predicted effect of the model after application of the improvements proposed herein.
As shown in the above table, the classification of 3D-SE ResNet10 on the data set herein works best. Is superior to both C3D Net and SE-C3D Net, and is superior to 3D residual neural networks without the introduction of attention mechanism. This illustrates: (1) in the case of using the 3D convolution kernel as well, the residual learning framework still performs better on the present data set, which is the same as the case of using the 2D convolution kernel as well, indicating that the advantage of the residual network framework on the present data set is universal; (2) the attention mechanism or the residual network framework can be introduced independently to achieve better effect on the data set; (3) the combined attention mechanism and residual network framework works better than either of the two alone.
In addition, as can be seen from comparing table 2 and table 3, under the condition that the network frameworks are similar, the classification effect of the data set used by the 3D neural network in the present invention is better than that of the 2D neural network, which fully demonstrates that the 3D convolution kernel can fully utilize the continuous features and spatial information in the medical image sequence, and achieve a better feature extraction effect than that of the 2D convolution kernel.
In conclusion, the three methods of using the 3D convolution kernel, introducing the attention mechanism and introducing the residual network framework are combined, so that the classification effect of the model on the data set used by the method can be improved to a greater extent.
The embodiment of the present application further provides a brain medical image classification apparatus, as shown in fig. 5, the brain medical image classification apparatus 400 includes:
an image acquisition module 401, configured to acquire a brain medical image sequence to be identified;
a preprocessing module 402, configured to preprocess the brain medical image sequence to obtain a brain medical image sequence with normalized data;
a classification module 403, configured to input the brain medical image sequence with normalized data into a trained image classification model, so as to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
The brain medical image classification method and device provided by the invention creatively and combinatively use two algorithms of a median filtering denoising algorithm and a histogram equalization algorithm as experimental data preprocessing algorithms, provide a 3D convolutional neural network based on a residual network framework, and introduce an attention mechanism (hereinafter abbreviated as SE-3D ResNet), wherein the algorithm designs a coping scheme of gradient disappearance based on a traditional convolutional neural network algorithm, integrates the attention mechanism to inhibit invalid information, improves the expression of effective information, accurately utilizes the spatial information of a stroke focus in a sequence to improve the feature extraction effect, and enables a model to distinguish whether a given brain medical image (sequence) is the stroke medical image (sequence) with higher accuracy. Experimental data show that the classification effect of the model can be improved to a greater extent by combining three methods of using a 3D convolution kernel, introducing an attention mechanism and introducing a residual network framework.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention.
Claims (8)
1. A brain medical image classification method is characterized by comprising the following steps:
acquiring a brain medical image sequence to be identified;
preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data;
inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
2. The method of claim 1, wherein the step of inputting the data-normalized brain medical image sequence into a trained image classification model to obtain the classification result of the brain medical image sequence comprises:
inputting the brain medical image sequence with the normalized data into the input module for convolution processing to obtain a primary effective characteristic diagram;
inputting the preliminary effective feature map into the residual error learning module for processing to obtain a key feature map;
and inputting the key feature map into the output module for classification processing to obtain a classification result of the brain medical image sequence.
3. The method for classifying medical brain images according to claim 2, wherein:
the residual learning module comprises 4 residual modules combined with an attention mechanism;
each residual module comprises a first convolution layer, a first BN layer, a first ReLU active layer, a second convolution layer, a second BN layer, an SE module and a second ReLU active layer which are connected in sequence;
the output of the SE module is added with the original information of the input module through jump connection, and then is used as the output of the SE-Res module through a ReLU activation function.
4. The method for classifying medical brain images according to claim 3, wherein the step of inputting the preliminary valid feature map into the residual learning module for processing to obtain a key feature map comprises:
processing the preliminary effective characteristic diagram x input into the residual error module to obtain a key characteristic diagram U by using the following formulac：
Uc＝SE(δ(Conv3×3×3(L(B(Conv3×3×3(x)))))+x
SE(in)＝σ(W2·δ(W1·Fsq(in)))
Where x represents the feature map of the most input residual module, B () represents the Batch Normalization operation, δ () represents the Leaky ReLU activation function, in represents the feature map of the input SE module, FsqFunction representation globalCarrying out average pooling operation; w1 and W2 represent parameters of two fully connected layers, respectively; σ (.) denotes Sigmoid activation function.
5. The method of claim 1, wherein the pre-processing the brain medical image sequence to obtain a data-normalized brain medical image sequence comprises:
reducing the original data of 512 × 512 pixels into a gray image of 224 × 224 pixels by using a nearest neighbor interpolation method;
processing salt-pepper noise and speckle noise in the brain medical image by using median filtering and denoising, so that the low-density shadow of the target region becomes a more continuous low-density shadow surface;
and improving the contrast of the brain medical image by using histogram equalization to obtain a brain medical image sequence with data normalization.
6. The method for classifying brain medical images according to claim 5, wherein improving the contrast of the brain medical images by histogram equalization comprises:
calculating the kth gray level r in a brain medical image by using the following formulakProbability of occurrence:
where n represents the total number of pixels in the image, L represents the number of gray levels in the image, nkRepresenting a grey level of rkThe number of pixels of (a); r iskIs the k-th gray level, p, in the brain medical imagek(rk) Representing the kth gray level r in the medical image of the brainkThe probability of occurrence.
Calculating the transformed gray level s of each pixel using the following histogram equalized discrete gray level transformation function formulak：
Wherein s iskThe gray level of each pixel after equalization is between 0 and 1, rjIs the j-th gray level, pr(rj) Is rjProbability of occurrence of gray scale, njTo a gray level of rjThe number of pixels of (a);
the gray level in the original image is rkEach pixel of (a) is mapped to a new image after histogram equalization as skTo obtain the brain medical image sequence with data normalization.
7. The method for classifying medical brain images according to claim 6, wherein the step of training the image classification model comprises:
acquiring a brain medical image CT sequence of a cerebral apoplexy patient;
preprocessing the brain medical image CT sequence of the cerebral apoplexy patient to obtain a data set containing the brain medical image CT sequence of the cerebral apoplexy patient with data normalization;
dividing the data set into a training set and a validation set;
and training the image classification model by using the training set, and selecting parameters of the entity recognition model by using the verification set to obtain the trained image classification model.
8. A brain medical image classification device, comprising:
the image acquisition module is used for acquiring a brain medical image sequence to be identified;
the preprocessing module is used for preprocessing the brain medical image sequence to obtain a brain medical image sequence with normalized data;
the classification module is used for inputting the brain medical image sequence with the normalized data into a trained image classification model to obtain a classification result of the brain medical image sequence; the image classification model is an SE-3D ResNet model and comprises an input module, a residual error learning module and an output module.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210021393.9A CN114359637A (en) | 2022-01-10 | 2022-01-10 | Brain medical image classification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210021393.9A CN114359637A (en) | 2022-01-10 | 2022-01-10 | Brain medical image classification method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114359637A true CN114359637A (en) | 2022-04-15 |
Family
ID=81109338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210021393.9A Pending CN114359637A (en) | 2022-01-10 | 2022-01-10 | Brain medical image classification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114359637A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115578593A (en) * | 2022-10-19 | 2023-01-06 | 北京建筑大学 | Domain adaptation method using residual attention module |
| CN117253092A (en) * | 2023-10-10 | 2023-12-19 | 江南大学 | Machine vision-based bin video classification and identification method and system |
| CN117392672A (en) * | 2023-12-11 | 2024-01-12 | 季华实验室 | Method for acquiring flow cell classification model, classification method and related equipment |
-
2022
- 2022-01-10 CN CN202210021393.9A patent/CN114359637A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115578593A (en) * | 2022-10-19 | 2023-01-06 | 北京建筑大学 | Domain adaptation method using residual attention module |
| CN117253092A (en) * | 2023-10-10 | 2023-12-19 | 江南大学 | Machine vision-based bin video classification and identification method and system |
| CN117392672A (en) * | 2023-12-11 | 2024-01-12 | 季华实验室 | Method for acquiring flow cell classification model, classification method and related equipment |
| CN117392672B (en) * | 2023-12-11 | 2024-03-19 | 季华实验室 | Method for acquiring flow cell classification model, classification method and related equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110930416B (en) | MRI image prostate segmentation method based on U-shaped network | |
| CN114359637A (en) | Brain medical image classification method and device | |
| CN109902558B (en) | CNN-LSTM-based human health deep learning prediction method | |
| CN111798462B (en) | Automatic delineation method of nasopharyngeal carcinoma radiotherapy target area based on CT image | |
| CN110349162B (en) | Retina macular edema multi-lesion image segmentation method | |
| CN112150476B (en) | Coronary artery sequence blood vessel segmentation method based on space-time discriminant feature learning | |
| CN110503630A (en) | A kind of cerebral hemorrhage classification, positioning and prediction technique based on three dimensional depth learning model | |
| US11222425B2 (en) | Organs at risk auto-contouring system and methods | |
| CN112488234A (en) | End-to-end histopathology image classification method based on attention pooling | |
| CN110910317B (en) | Tongue image enhancement method | |
| CN113592794B (en) | Spine graph segmentation method of 2D convolutional neural network based on mixed attention mechanism | |
| CN107506766B (en) | Image segmentation method and device | |
| CN113554665A (en) | Blood vessel segmentation method and device | |
| CN110211671B (en) | Thresholding method based on weight distribution | |
| CN114723669A (en) | Liver tumor two-point five-dimensional deep learning segmentation algorithm based on context information perception | |
| CN113989551A (en) | Alzheimer disease classification method based on improved ResNet network | |
| CN115082493A (en) | 3D (three-dimensional) atrial image segmentation method and system based on shape-guided dual consistency | |
| CN113139931B (en) | Thyroid section image classification model training method and device | |
| CN113436128B (en) | Dual-discriminator multi-mode MR image fusion method, system and terminal | |
| CN113269774B (en) | Parkinson disease classification and lesion region labeling method of MRI (magnetic resonance imaging) image | |
| CN111563858B (en) | Denoising method of human embryo heart ultrasonic image based on depth convolution neural network | |
| CN113012164A (en) | U-Net kidney tumor image segmentation method and device based on inter-polymeric layer information and storage medium | |
| CN113077891A (en) | Big data disease diagnosis system based on algorithm, block chain and medical image | |
| CN111667469A (en) | Lung disease classification method, device and equipment | |
| CN116563554A (en) | Low-dose CT image denoising method based on hybrid characterization learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |