Brain tumor image segmentation method based on high-order attention mechanism

Abstract

The invention discloses a brain tumor image segmentation method based on a high-order attention mechanism, and the method comprises the steps: carrying out the data preprocessing of a brain tumor MRI (Magnetic Resonance Imaging) image data set, obtaining a preprocessed brain tumor image data set, dividing the brain tumor image data set into a training data set and a test data set according to a proportion, the improved HLNet model is trained and verified through a training data set and a test data set, a trained brain tumor segmentation model is finally obtained, a brain tumor image is segmented, and Dice coefficients of an enhanced tumor (ET), a tumor core (TC) and a whole tumor (WT) are obtained. According to the invention, the brain tumor segmentation image can be rapidly and accurately obtained.

Description

Technical Field

[0001] The invention relates to a method for brain tumor image recognition, and specifically to a method for brain tumor image recognition based on deep learning. Technical Background

[0002] Accurate segmentation of brain tumors is one of the important tasks in medical image analysis and clinical diagnosis, especially in the planning of brain tumor surgery and the formulation of treatment plans. At present, traditional segmentation methods mainly rely on manual annotation and expert experience, and radiologists need to manually outline the tumor area, which is not only time-consuming and labor-intensive, but also easily affected by the subjective factors of the annotator, resulting in certain deviations in the segmentation results.

[0003] With the development of deep learning technology, it has demonstrated powerful automation and efficiency in the field of medical image segmentation. In particular, segmentation methods based on convolutional neural networks have achieved remarkable results in many fields. However, for brain tumor MRI image segmentation, due to the structural complexity and richness of detailed information of different tumor areas in the image data (such as enhanced tumors, necrotic areas, peripheral edema, etc.), the existing deep learning models still have the problem of insufficient capture of local detailed information, resulting in the accuracy and precision of the segmentation results being difficult to meet clinical needs.

[0004] Chinese patent application CN118446990A discloses a brain tumor segmentation method based on 3D-Unet network model optimization. This method is applied to brain tumor magnetic resonance imaging (MRI). In view of the problem that the network has a slow convergence speed and contains a large amount of redundant information, resulting in poor segmentation results, this method introduces the VIT (Vision Transformer) module on 3D-UNet and uses the unique position encoding in the model to improve the network convergence speed; at the same time, an attention gate mechanism and a multi-scale residual module are added to the model, and the attention gate mechanism is introduced to suppress the activation of features in irrelevant areas to improve the model sensitivity and accuracy, ensure the importance of segmentation features, avoid interference from redundant information features, and improve the accuracy of brain tumor segmentation. However, this technology only extracts features from images at a single scale, and cannot extract feature information of images from multiple scales, resulting in insufficient accuracy of processed images. Summary of the invention

[0005] In order to effectively solve the above technical problems, the present invention provides a brain tumor image segmentation method based on a high-order attention mechanism. The high-order attention mechanism is used to extract multi-scale features from brain tumor MRI images, making full use of the detailed information in the image, greatly improving the accuracy and stability of segmentation.

[0006] In order to achieve the above object, the present invention proposes a brain tumor image segmentation method based on a high-order attention mechanism, which specifically includes the following steps:

[0007] Step S1, performing data preprocessing on the MRI image data set, including normalization processing and cropping operations, to obtain a preprocessed MRI image data set;

[0008] Step S2, dividing the data-enhanced brain tumor slice image dataset into a training dataset and a test dataset in a ratio of 8:2;

[0009] Step S3, construct a high-order attention model based on U-Net++, integrate the high-order attention module in the encoder, decoder and skip connection of U-Net++ to obtain an improved HLNet model;

[0010] Step S4, inputting the training data set into the improved model for training, and continuously optimizing the Dice coefficient of the model by adjusting the deep learning hyperparameters until the Dice coefficient of the model converges, thereby obtaining a trained brain tumor segmentation model;

[0011] Step S5: Segment the brain tumor image using the trained brain tumor segmentation model to obtain Dice coefficients of the enhanced tumor (ET), the tumor core (TC), and the whole tumor (WT).

[0012] Preferably, the MRI image data set includes three segmented regions, namely enhanced tumor (ET), tumor core (TC) and whole tumor (WT), and covers a variety of image sequences such as T1, T1ce, T2 and FLAIR images.

[0013] Optionally, the data enhancement includes: performing standardization, cropping, and slicing on each of the brain tumor images in sequence.

[0014] Optionally, the standardization processing includes: performing percentile-based range restrictions on each of the different modalities (such as FLAIR, T1, T1ce, T2) of the brain tumor image, and using a z-score standardization method so that each modality of the image has a zero mean and a unit standard deviation, thereby obtaining standardized processing data;

[0015] Optionally, the cropping includes: locking the central area of ​​each of the brain tumor images, cropping the image into an area of ​​160×160 pixels to ensure that the processed image has a uniform input size, and obtaining cropped data;

[0016] Optionally, the slice processing includes: for each of the cropped images, processing each slice in turn, removing slices without lesions, ensuring that the training data only contains image data containing the tumor area, and the enhanced MRI image data set.

[0017] Preferably, S3 specifically includes:

[0018] Build a basic U-Net++ model, determine its basic architecture and the initial settings of each layer component, including convolution layer parameters, jump connection mode, etc. Integrate the advanced attention mechanism module into the U-Net++ model in a specific way, that is, introduce the module in the encoder, decoder and jump connection respectively, clarify its connection mode and mechanism in each part, and ensure that it can effectively extract features at different levels and perform information fusion. At the same time, combined with the actual data characteristics and task requirements, optimize the model's loss function, adjust the weight ratio of binary cross entropy loss (BCE) and Dice loss, so that the model performs better in processing category imbalance data (such as the difference in the number of tumor and background pixels) and fuzzy boundary problems, and finally obtain the improved HLNet model.

[0019] Preferably, the expression of the Loss loss function is:

[0020] Loss function:

[0021] Where N is the number of samples, y i It's the real label. is the predicted output of the model, |X∩Y| represents the number of elements in the intersection between X and Y, and |X| and |Y| represent the number of elements in X and Y.

[0022] Preferably, S4 is specifically:

[0023] The training data set is input into the improved model for training, and the Dice coefficient of the model is continuously optimized by adjusting the deep learning hyperparameters until the Dice coefficient of the model converges, thereby obtaining a trained brain tumor segmentation model.

[0024] Preferably, the expression for calculating the model accuracy is:

[0025] Accuracy:

[0026] In the formula, TP (True Positives) represents the number of positive samples correctly predicted as positive; FP (False Positives) represents the number of negative samples incorrectly predicted as positive; FN (False Negatives) represents the number of positive samples incorrectly predicted as negative.

[0027] Preferably, the deep learning hyperparameters include: training batch size, number of iterations, initial learning rate, and HLNet model parameters.

[0028] Preferably, S5 specifically includes: segmenting the brain tumor image by using the improved model, calculating the Dice coefficient according to the output result, and selecting the segmentation result with the highest Dice coefficient as the final segmentation result of the model.

[0029] Preferably

[0030] The advantages of the present invention are:

[0031] 1) The model in this paper extracts features from each pixel in the brain tumor image without manual operation, thus avoiding the influence of human factors on the segmentation results and improving the segmentation accuracy.

[0032] 2) Compared with traditional technologies, the present invention is not affected by human factors and can efficiently and accurately perform brain tumor image segmentation with almost no additional cost or human intervention.

[0033] 3) Compared with the existing technology, the present invention combines multimodal information and introduces a high-order attention mechanism to improve the segmentation accuracy of complex brain tumor MRI images. It can output accurate results in real time and assist doctors in accurate diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. It is obvious that the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 This is a structural diagram of the high-order attention module of the present invention;

[0037] Figure 3 This is a structural diagram of the improved HLNet model of the present invention;

[0038] Figure 4 This is an example diagram of the brain tumor image recognition result of the present invention.

[0039] Figure 5 This is the improved HLNet model identification step diagram of the present invention DETAILED DESCRIPTION

[0040] The present invention is described clearly and completely below in conjunction with the accompanying drawings. A person skilled in the art will be able to implement the present invention based on these descriptions. Before describing the present invention in conjunction with the accompanying drawings, it should be particularly noted that:

[0041] The technical solutions and technical features provided in each part of the present invention, including the following description, may be combined with each other if there is no conflict.

[0042] In addition, the embodiments of the present invention involved in the following description are generally only a part of the embodiments of the present invention, rather than all the embodiments. Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0043] About the terms and units in the present invention: The term "comprises" and any variations thereof in the description and claims of the present invention and related parts are intended to cover non-exclusive inclusions.

[0044] Example 1

[0045] Reference Figure 1 As shown, the present invention proposes a brain tumor image segmentation method based on a high-order attention mechanism, and the method specifically includes the following steps:

[0046] Step S1, performing data preprocessing on the MRI image data set, including normalization processing and cropping operations, to obtain a preprocessed MRI image data set.

[0047] This step is used to convert the three-dimensional data set into a two-dimensional data set to prevent the poor model fitting effect due to the large data set and further improve the generalization ability of the model.

[0048] In this embodiment, the data enhancement in step S1 includes: performing standardization, cropping, and slicing on each of the brain tumor images in sequence.

[0049] In this embodiment, the data enhancement operation includes: performing standardization, cropping, and slicing on each of the brain tumor images in sequence.

[0050] In this embodiment, the standardization processing includes: performing percentile-based range restrictions on each of the different modalities of the brain tumor image (such as FLAIR, T1, T1ce, T2), and using a z-score standardization method so that each modality of the image has a zero mean and a unit standard deviation to obtain standardized processed data.

[0051] In this embodiment, the cropping includes: locking the central area of ​​each of the brain tumor images, cropping the image into an area of ​​160×160 pixels, so as to ensure that the processed image has a uniform input size, and obtaining cropped data.

[0052] In this embodiment, the slice processing includes: for each of the cropped images, processing each slice in turn, removing slices without lesions, ensuring that the training data only contains image data containing tumor areas, and obtaining an enhanced MRI image dataset. The enhanced brain tumor image dataset contains a total of 22,142 brain tumor slice images.

[0053] Step S2: divide the data-enhanced brain tumor slice image dataset into a training dataset and a test dataset in a ratio of 8:2. The training dataset includes 18,923 brain tumor slice images, and the test dataset includes 3,219 brain tumor slice images.

[0054] Step S3, construct a high-order attention model based on U-Net++, integrate the high-order attention module in the encoder, decoder and skip connection of U-Net++ to obtain an improved HLNet model;

[0055] The U-Net++ model extracts richer features by connecting different levels of output and high-order attention modules. The structure of the high-order attention module is as follows: Figure 2 As shown in the figure, high-order attention modeling is performed on each layer of input feature map to capture detailed information of different levels and dimensions in the input feature map. In the forward propagation process, multiple attention maps are first calculated through multiple convolution sub-modules (each sub-module contains a convolution layer and a Sigmoid activation function). Then, these attention maps are combined according to the level to generate a comprehensive high-order attention representation. Finally, this high-order attention representation is multiplied with the input feature map to obtain the final output, and the improved HLNet model is obtained.

[0056] The structure of the improved HLNet model is as follows Figure 3 As shown in the figure, Resblock is a convolution layer for convolution processing, Downsampling with a downward arrow includes the downsampling operation, and Upsampling with an upward arrow includes the upsampling operation. HL represents the high-order attention module, which can enhance the model's ability to focus on key areas. It can help the model capture important features of the tumor area.

[0057] Step S4: input the training data set into the improved model for training, and continuously optimize the Dice coefficient of the model by adjusting the deep learning hyperparameters until the Dice coefficient of the model converges to obtain a trained brain tumor segmentation model.

[0058] Preferably, the expression for calculating the model accuracy is:

[0059] Accuracy:

[0060] In the formula, TP (True Positives) represents the number of positive samples correctly predicted as positive; FP (False Positives) represents the number of negative samples incorrectly predicted as positive; FN (False Negatives) represents the number of positive samples incorrectly predicted as negative.

[0061] The deep learning hyperparameters include: training batch size, number of iterations, initial learning rate, and HLNet model parameters.

[0062] Step S5: Segment the brain tumor image using the trained brain tumor segmentation model to obtain Dice coefficients of the enhanced tumor (ET), the tumor core (TC), and the whole tumor (WT).

[0063] The improved HLNet model segments the brain tumor image, calculates the Dice coefficients of the enhanced tumor (ET), tumor core (TC) and whole tumor (WT), and selects the result with the highest Dice coefficient as the final segmentation result. Given brain tumor images of enhanced tumor (ET), tumor core (TC) and whole tumor (WT), respectively, Figure 4 As shown in Figure 1, the improved HLNet model segmentation results are compared with the manual segmentation results. The improved HLNet model segmentation is accurate. In the figure, red represents enhanced tumor (ET), yellow represents tumor core (TC), and green represents whole tumor (WT). The overall segmentation steps of the brain tumor image segmentation model are as follows: Figure 5 Shown

[0064] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A brain tumor image segmentation method based on a high-order attention mechanism, characterized in that: The specific steps include: Step S1, performing data preprocessing on the MRI image data set, including normalization processing and cropping operations, to obtain a preprocessed MRI image data set; Step S2, dividing the data-enhanced brain tumor slice image dataset into a training dataset and a test dataset in a ratio of 8:2; Step S3, construct a high-order attention model based on U-Net++, integrate the high-order attention module in the encoder, decoder and skip connection of U-Net++ to obtain an improved HLNet model; Step S4, inputting the training data set into the improved model for training, and continuously optimizing the Dice coefficient of the model by adjusting the deep learning hyperparameters until the Dice coefficient of the model converges, thereby obtaining a trained brain tumor segmentation model; Step S5: Segment the brain tumor image using the trained brain tumor segmentation model to obtain Dice coefficients of the enhanced tumor (ET), the tumor core (TC), and the whole tumor (WT).

2. The brain tumor image segmentation method based on high-order attention mechanism according to claim 1, characterized in that: The MRI image dataset includes three segmented regions: enhanced tumor (ET), tumor core (TC) and whole tumor (WT), covering multiple image sequences such as T1, T1ce, T2 and FLAIR images, and 22,142 brain tumor slice images.

3. The brain tumor image segmentation method based on high-order attention mechanism according to claim 1, characterized in that: The data enhancement operation includes: performing standardization, cropping, and slicing on each of the brain tumor images in sequence.

4. The brain tumor image segmentation method based on high-order attention mechanism according to claim 1, characterized in that: The S3 is specifically: Build a basic U-Net++ model, determine its basic architecture and the initial settings of each layer component, including convolution layer parameters, jump connection mode, etc. Integrate the advanced attention mechanism module into the U-Net++ model in a specific way, that is, introduce the module in the encoder, decoder and jump connection respectively, clarify its connection mode and mechanism in each part, and ensure that it can effectively extract features at different levels and perform information fusion. At the same time, combined with the actual data characteristics and task requirements, optimize the model's loss function, adjust the weight ratio of binary cross entropy loss (BCE) and Dice loss, so that the model performs better in processing category imbalance data (such as the difference in the number of tumor and background pixels) and fuzzy boundary problems, and finally obtain the improved HLNet model.

5. The brain tumor image segmentation method based on high-order attention mechanism according to claim 4, characterized in that: The expression of the Loss loss function is: Loss function: Where N is the number of samples, y i It's the real label. is the predicted output of the model, |X ∩ Y| represents the number of elements in the intersection between X and Y, and |X| and |Y| represent the number of elements in X and Y.

6. The brain tumor image segmentation method based on high-order attention mechanism according to claim 1, characterized in that: The S4 is specifically: The training data set is input into the improved model for training, and the Dice coefficient of the model is continuously optimized by adjusting the deep learning hyperparameters until the Dice coefficient of the model converges, thereby obtaining a trained brain tumor segmentation model.

7. The brain tumor image segmentation method based on high-order attention mechanism according to claim 5 is characterized in that The expression for calculating the accuracy of the model is: Accuracy: In the formula, TP (True Positives) represents the number of positive samples correctly predicted as positive; FP (False Positives) represents the number of negative samples incorrectly predicted as positive; FN (False Negatives) represents the number of positive samples incorrectly predicted as negative. The deep learning hyperparameters include: training batch size, number of iterations, initial learning rate, and HLNet model parameters.

8. The brain tumor image segmentation method based on high-order attention mechanism according to claim 1, characterized in that: The S5 is specifically: The brain tumor image is segmented by the improved model, the Dice coefficient is calculated according to the output result, and the segmentation result with the highest Dice coefficient is selected as the final segmentation result of the model.

Images