Intelligent detection method for brain tumor focus area

Abstract

The invention discloses an intelligent detection method for a brain tumor focus area, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring a multi-modal image; the multi-modal image is an MRI (Magnetic Resonance Imaging) image generated according to T1 weighting, T2 weighting and an FLAIR sequence; the multi-modal image is preprocessed, and a normalized image is obtained; constructing a focus detection model; the focus detection model adopts a double-branch network structure and comprises a global branch, a local branch and a weighted fusion module; the global branch and the local branch are both connected with the weighted fusion module; the weight fusion module adopts a dynamic weight fusion strategy; inputting the normalized image into the focus detection model to generate a preliminary focus segmentation result; and performing post-processing optimization on the preliminary focus segmentation result to generate a final focus positioning result. According to the invention, the detection precision of the brain tumor lesion area can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an intelligent detection method for brain tumor lesion areas. Background Art

[0002] Brain tumors are serious neurological diseases that threaten human health. Early and accurate diagnosis is crucial for developing treatment plans and assessing prognosis. Currently, the diagnosis of brain tumors relies primarily on magnetic resonance imaging (MRI), which uses multimodal imaging techniques such as T1-weighted, T2-weighted, and FLAIR sequences to obtain information on the tumor's location, size, and morphology. However, traditional manual image reading relies heavily on the physician's experience and subjective judgment, resulting in low efficiency, prone to missed or misdiagnosed diagnoses, and difficulty in meeting the needs of analyzing large-scale imaging data.

[0003] In recent years, automated detection methods based on deep learning have been gradually applied to the field of medical imaging, but existing technologies still have shortcomings in brain tumor lesion detection. For example, some methods have limited ability to fuse and analyze multimodal images, failing to fully utilize the complementary information of different imaging sequences; some methods lack accuracy in lesion boundary identification, which can easily lead to incomplete lesion segmentation or the occurrence of false-positive areas; and some methods have high computational complexity, making it difficult to meet the needs of clinical real-time detection. Therefore, developing a high-precision, low-resource intelligent detection method for brain tumor lesion regions that fully utilizes multimodal imaging information has important clinical application value. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent detection method for brain tumor lesion areas, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for intelligently detecting brain tumor lesion areas, comprising:

[0007] Acquire multimodal images; the multimodal images are MRI images generated based on T1-weighted, T2-weighted, and FLAIR sequences;

[0008] Preprocessing the multimodal image to obtain a normalized image;

[0009] Constructing a lesion detection model; the lesion detection model adopts a dual-branch network structure, including a global branch, a local branch and a weighted fusion module; the global branch and the local branch are both connected to the weighted fusion module; the weighted fusion module adopts a dynamic weight fusion strategy;

[0010] Inputting the normalized image into the lesion detection model to generate a preliminary lesion segmentation result;

[0011] The preliminary lesion segmentation result is post-processed and optimized to generate a final lesion localization result.

[0012] Optionally, preprocessing the multimodal image to obtain a normalized image specifically includes:

[0013] The multimodal images are registered, skull stripped and grayscale normalized to obtain a normalized image.

[0014] Optionally, the process of constructing the lesion detection model includes:

[0015] Obtaining a training data set; the training data set includes training images and corresponding label data;

[0016] Constructing a pre-trained network based on a dual-branch network structure; in the pre-trained network, the global branch is used to extract global features of the image; the local branch is used to extract local features of the image; the weighted fusion module is used to perform weighted fusion of the global features and the local features;

[0017] The training data set is input into the pre-trained network, and iterative training is performed according to the set loss function and gradient descent strategy, and the network that meets the accuracy evaluation threshold is determined as the lesion detection model.

[0018] Optionally, the set loss function adopts Huber Loss loss function and cross entropy loss function.

[0019] Optionally, both the global branch and the local branch are constructed using an improved U-Net network; wherein, the improved U-Net network used by the global branch is a multi-layer 3D CNN encoder and decoder structure; the improved U-Net network used by the local branch is a spatial attention mechanism introduced into the U-Net architecture.

[0020] Optionally, inputting the normalized image into the lesion detection model to generate a preliminary lesion segmentation result specifically includes:

[0021] The normalized image is simultaneously input into the global branch and the local branch. In the global branch, a 5-layer 3D CNN encoder and decoder structure is used to extract features to obtain global features. A spatial attention module is introduced in the decoder stage of the local branch to perform weighted enhancement on key features of the lesion area to obtain local features.

[0022] The global features and the local features are input into the weighted fusion module for dynamic weighted fusion to generate a preliminary lesion segmentation result.

[0023] Optionally, the post-processing optimization of the preliminary lesion segmentation result to generate a final lesion localization result specifically includes:

[0024] The preliminary lesion segmentation result is subjected to edge smoothing processing using a fully connected conditional random field, and false positive areas are removed through connected domain analysis to generate a final lesion localization result; the final lesion localization result includes the lesion volume and position coordinates.

[0025] Optionally, the false positive area is the area of ​​the segmented result that is less than 10mm 2 area.

[0026] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0027] The present invention discloses an intelligent method for detecting brain tumor lesion areas. The method comprises acquiring multimodal images; the multimodal images are MRI images generated based on T1-weighted, T2-weighted, and FLAIR sequences; preprocessing the multimodal images to obtain normalized images; constructing a lesion detection model; the lesion detection model adopts a dual-branch network structure, including a global branch, a local branch, and a weighted fusion module; the global branch and the local branch are both connected to the weighted fusion module; the weighted fusion module adopts a dynamic weight fusion strategy; the normalized image is input into the lesion detection model to generate a preliminary lesion segmentation result; the preliminary lesion segmentation result is post-processed and optimized to generate a final lesion localization result. The present invention can improve the detection accuracy of brain tumor lesion areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, 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 any creative work.

[0029] Figure 1 Schematic diagram of the process of the intelligent detection method of brain tumor lesion area of ​​the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] The purpose of the present invention is to provide an intelligent detection method for brain tumor lesion areas, aiming to solve or improve at least one of the above-mentioned technical problems.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, the present invention provides an intelligent detection method for brain tumor lesion area, comprising:

[0034] A method for intelligently detecting brain tumor lesion areas, comprising:

[0035] Step 100: Acquire a multimodal image; the multimodal image is an MRI image generated according to T1-weighted, T2-weighted, and FLAIR sequences.

[0036] Step 200: Preprocess the multimodal image to obtain a normalized image.

[0037] Step 300: Construct a lesion detection model; the lesion detection model adopts a dual-branch network structure, including a global branch, a local branch and a weighted fusion module; the global branch and the local branch are both connected to the weighted fusion module; the weighted fusion module adopts a dynamic weight fusion strategy.

[0038] Step 400: Input the normalized image into the lesion detection model to generate a preliminary lesion segmentation result.

[0039] Step 500: performing post-processing optimization on the preliminary lesion segmentation result to generate a final lesion localization result.

[0040] As a specific implementation manner, the preprocessing of the multimodal image to obtain a normalized image specifically includes: performing registration, skull stripping and grayscale normalization on the multimodal image to obtain a normalized image.

[0041] As a specific implementation, the process of constructing the lesion detection model includes:

[0042] A training dataset is obtained; the training dataset includes training images and corresponding label data; a pre-trained network based on a dual-branch network structure is constructed; in the pre-trained network, the global branch is used to extract global features of the image; the local branch is used to extract local features of the image; the weighted fusion module is used to perform weighted fusion of the global features and the local features; the training dataset is input into the pre-trained network, and iterative training is performed according to a set loss function and gradient descent strategy. The network that meets the accuracy evaluation threshold is determined as the lesion detection model. The set loss function adopts the Huber Loss loss function and the cross entropy loss function.

[0043] As a specific implementation method, the global branch and the local branch are both constructed using an improved U-Net network; wherein, the improved U-Net network used by the global branch is a multi-layer 3DCNN encoder and decoder structure; the improved U-Net network used by the local branch is a spatial attention mechanism introduced into the U-Net architecture.

[0044] As a specific implementation, inputting the normalized image into the lesion detection model to generate a preliminary lesion segmentation result specifically includes:

[0045] The normalized image is simultaneously input into the global branch and the local branch. In the global branch, a 5-layer 3D CNN encoder and decoder structure is used to extract features to obtain global features. A spatial attention module is introduced into the decoder stage of the local branch to perform weighted enhancement on the key features of the lesion area to obtain local features. The global features and the local features are input into the weighted fusion module for dynamic weighted fusion to generate a preliminary lesion segmentation result.

[0046] As a specific implementation manner, the post-processing optimization of the preliminary lesion segmentation result to generate the final lesion localization result specifically includes:

[0047] The edge smoothing process of the preliminary lesion segmentation result is performed using a fully connected conditional random field, and false positive areas are removed through connected domain analysis to generate the final lesion localization result; the final lesion localization result includes the lesion volume and position coordinates. Among them, the false positive area is the area of ​​the segmented result that is less than 10mm 2 area.

[0048] Based on the above technical solution, the following embodiments are provided.

[0049] In this embodiment, MRI image data of glioma patients from a tertiary hospital are used as verification objects, and the intelligent detection method for brain tumor lesion areas proposed in the present invention is used to perform lesion segmentation and positioning.

[0050] First, data acquisition and preprocessing are performed.

[0051] 1) Data Collection

[0052] Multimodal image acquisition: MRI images of brain tumor patients were extracted from the hospital's PACS system, including T1-weighted (T1w), T2-weighted (T2w), and fluid-attenuated inversion recovery (FLAIR) sequences.

[0053] Modal function: T1w is used to clearly display the anatomical structure and assist in locating the lesion; T2w is used to highlight the edema area and assist in determining the tumor boundary; FLAIR is used to suppress cerebrospinal fluid signals and enhance the contrast between lesions and normal tissues.

[0054] Data screening criteria: The images must include the complete brain area without severe motion artifacts or metal artifacts; the cases must be pathologically confirmed as brain tumors (such as gliomas, meningiomas, etc.).

[0055] 2) Image preprocessing

[0056] Skull Stripping: Deep learning-based skull stripping tools (such as DeepBrain) are used to remove the skull and scalp tissue, leaving only the brain parenchyma. This aims to reduce the interference of non-brain tissue on lesion detection and improve model training efficiency.

[0057] Registration: Rigid registration (6 degrees of freedom) of T1w, T2w, and FLAIR images was performed using the FLIRT tool in FSL software to ensure spatial alignment of the multimodal images. Mutual information was used as the similarity metric.

[0058] Grayscale normalization (Intensity Normalization): Z-score normalization is performed on the images of each modality to eliminate the grayscale differences caused by different devices and scanning parameters.

[0059] Data augmentation: Perform random rotation (±10°), flipping (horizontally / vertically), and elastic deformation on the training set to expand the data scale and improve the model generalization ability.

[0060] Then, a lesion detection model is constructed.

[0061] Network structure design:

[0062] Global Branch: A five-layer 3D convolutional neural network (3D CNN) encoder-decoder architecture is employed. In the encoder stage, each layer consists of two 3×3×3 convolutional layers (stride = 1, padding = 1), batch normalization (BatchNorm), and ReLU activation, followed by a 2×2×2 max pooling layer (stride = 2). In the decoder stage, each layer consists of a 2×2×2 transposed convolution (stride = 2) for upsampling, a skip connection (SkipConnection) to the feature map of the symmetric encoder layer, followed by two 3×3×3 convolutional layers. At the output stage, a global feature map of the image is generated (1 / 16 the size of the original image).

[0063] Local Branch: The spatial attention module (SAM) is introduced in the U-Net decoder stage to enhance the local features of the lesion area.

[0064] The Spatial Attention Module (SAM) is implemented by performing global average pooling on the decoder feature map to generate a channel attention vector. A spatial attention weight map is generated through 1×1×1 convolution and sigmoid activation. The weight map is element-wise multiplied with the original feature map to achieve feature weighting. At the output stage, a local feature map of the lesion region is generated (with the same size as the global branch).

[0065] Weighted Fusion Module:

[0066] Dynamic weight generation: Global features and local features are weighted through a lightweight network (such as a 2-layer fully connected layer), and the weight coefficients are adaptively learned from the training data.

[0067] Then, model training is performed.

[0068] The training dataset contains images of 40 patients, each with lesion areas annotated by a professional physician (GroundTruth). The data is divided into a training set (32 cases) and a validation set (8 cases).

[0069] The Huber Loss function was used to reduce the impact of outliers on gradient updates, and the cross-entropy loss function was used to optimize the classification task (lesion / non-lesion). The Adam optimizer was further used, with the learning rate initialized to 0.0001 and decayed to the original 0.9 every 20 rounds.

[0070] Set the training termination condition: the Dice coefficient does not improve for 10 consecutive rounds on the validation set, or the maximum number of training rounds (500 rounds) is reached.

[0071] Finally, lesion detection and post-processing optimization are performed.

[0072] Initial lesion segmentation:

[0073] The preprocessed image (size 256×256×16) is fed into both the global and local branches for feature extraction. The global branch extracts global image features using a five-layer 3D CNN encoder-decoder. The local branch uses SAM (Sensor-Amplifier) ​​to focus on lesion boundaries during the decoder phase, generating local feature maps. Dynamic weighted fusion is then used to feed the global and local features into a weighted fusion module to generate a preliminary segmentation result (with the same size as the input image).

[0074] Post-processing optimization:

[0075] Edge smoothing (Dense CRF) uses fully connected conditional random fields (DenseCRF) to optimize the edges of the initial segmentation results. Connected component analysis (Connected Component Analysis) is used to count the connected areas of the segmentation results and remove areas smaller than 10mm. 3 The area (assuming a voxel resolution of 1×1×1mm 3 ).

[0076] After the above experimental process, we can get:

[0077] Evaluation indicators:

[0078] Dice coefficient: measures the degree of overlap between the segmentation result and the ground truth;

[0079] Sensitivity: the proportion of correctly detected lesions;

[0080] False positive rate (FPR): The proportion of non-lesion areas that are mistakenly detected as lesions.

[0081] Comparative experiment:

[0082] Traditional U-Net: Dice coefficient = 0.785, sensitivity = 82.3%, FPR = 12.1%;

[0083] The method of the present invention has the following characteristics: Dice coefficient = 0.912, sensitivity = 88.7%, FPR = 6.3%.

[0084] Therefore, the solution described in this embodiment achieves a single image processing time of 0.78 seconds for real-time detection, meeting the needs of emergency scenarios. Furthermore, the sensitivity for lesions with a diameter of less than 5 mm is increased by 15.4%, enabling the detection of even tiny lesions. Finally, the generated 3D lesion model can assist physicians in surgical planning. Through the above steps, the method of the present invention achieves high-precision, real-time detection of brain tumor lesions, providing reliable technical support for clinical diagnosis.

[0085] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0086] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for intelligent detection of brain tumor lesion areas, characterized in that: include: Acquire multimodal images; the multimodal images are MRI images generated based on T1-weighted, T2-weighted, and FLAIR sequences; Preprocessing the multimodal image to obtain a normalized image; Constructing a lesion detection model; the lesion detection model adopts a dual-branch network structure, including a global branch, a local branch and a weighted fusion module; the global branch and the local branch are both connected to the weighted fusion module; the weighted fusion module adopts a dynamic weight fusion strategy; Inputting the normalized image into the lesion detection model to generate a preliminary lesion segmentation result; The preliminary lesion segmentation result is post-processed and optimized to generate a final lesion localization result.

2. The intelligent detection method for brain tumor lesion area according to claim 1, characterized in that: The preprocessing of the multimodal image to obtain a normalized image specifically includes: The multimodal images are registered, skull stripped and grayscale normalized to obtain a normalized image.

3. The intelligent detection method for brain tumor lesion area according to claim 1, characterized in that: The construction process of the lesion detection model includes: Obtaining a training data set; the training data set includes training images and corresponding label data; Constructing a pre-trained network based on a dual-branch network structure; in the pre-trained network, the global branch is used to extract global features of the image; the local branch is used to extract local features of the image; the weighted fusion module is used to perform weighted fusion of the global features and the local features; The training data set is input into the pre-trained network, and iterative training is performed according to the set loss function and gradient descent strategy, and the network that meets the accuracy evaluation threshold is determined as the lesion detection model.

4. The intelligent detection method for brain tumor lesion area according to claim 3, characterized in that: The loss function used in the setting is the Huber Loss loss function and the cross entropy loss function.

5. The intelligent detection method for brain tumor lesion area according to claim 1, characterized in that: Both the global branch and the local branch are constructed using an improved U-Net network; wherein, the improved U-Net network used by the global branch is a multi-layer 3D CNN encoder and decoder structure; the improved U-Net network used by the local branch is a spatial attention mechanism introduced into the U-Net architecture.

6. The intelligent detection method for brain tumor lesion area according to claim 1, characterized in that: Inputting the normalized image into the lesion detection model to generate a preliminary lesion segmentation result specifically includes: The normalized image is simultaneously input into the global branch and the local branch. In the global branch, a 5-layer 3D CNN encoder and decoder structure is used to extract features to obtain global features. A spatial attention module is introduced in the decoder stage of the local branch to perform weighted enhancement on key features of the lesion area to obtain local features. The global features and the local features are input into the weighted fusion module for dynamic weighted fusion to generate a preliminary lesion segmentation result.

7. The intelligent detection method for brain tumor lesion area according to claim 1, characterized in that: The post-processing optimization of the preliminary lesion segmentation result to generate the final lesion localization result specifically includes: The preliminary lesion segmentation result is subjected to edge smoothing processing using a fully connected conditional random field, and false positive areas are removed through connected domain analysis to generate a final lesion localization result; the final lesion localization result includes the lesion volume and position coordinates.

8. The intelligent detection method for brain tumor lesion area according to claim 7, characterized in that: The false positive area is the area of ​​the segmented result that is less than 10mm 2 area.

Images