A brain tumor recognition analysis system and method of a magnetic resonance image

Abstract

The application discloses a brain tumor recognition and analysis system and method of a magnetic resonance image. The system comprises an image acquisition module, an image preprocessing module, an ROI automatic segmentation module, a feature extraction module, a classification and discrimination module, an interaction and output module, has the advantages of high automation degree, high classification precision, suitability for clinical auxiliary diagnosis and the like, the method can be embedded into a PACS system, realizes automatic import, analysis and visual output of the MRI image, supports doctor interactive correction, and has wide clinical application effect.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence image recognition, more particularly to the field of medical image recognition analysis, and specifically to an automatic brain tumor recognition analysis system and method based on multi-parameter magnetic resonance imaging (MRI) images. BACKGROUND

[0002] Primary sinonasal malignancies are relatively rare in clinical practice, among which melanoma and lymphoma are the most common pathological types. Although both of them can present similar features such as high signal on T1WI and low signal on T2WI on imaging, their biological behavior, treatment plan and prognosis are completely different. Melanoma usually progresses rapidly and has a poor prognosis, requiring early surgical resection and adjuvant therapy; while lymphoma is sensitive to radiotherapy and chemotherapy, and the treatment strategy is completely different from surgery. Therefore, accurate differentiation of the two on the imaging level is of great significance for clinical diagnosis and individualized treatment.

[0003] Magnetic resonance imaging (MRI) as a non-invasive imaging method can provide anatomical and functional information of tumors, especially diffusion weighted imaging (DWI) and dynamic contrast enhanced imaging (DCE-MRI) have important value in tumor typing and benign and malignant judgment. DWI quantitatively reflects the cell density and water molecule diffusion degree through the apparent diffusion coefficient (ADC), while DCE-MRI reflects the angiogenesis and perfusion characteristics of the tumor, providing functional clues for distinguishing different types of tumors.

[0004] However, the existing technology still faces the following problems in the image analysis process:

[0005] The ROI delineation process relies on manual operation and has strong subjectivity and poor repeatability, especially when the tumor boundary is blurred or the heterogeneity is strong, which is easy to introduce errors;

[0006] Most image discrimination methods are based on experience or linear models, such as Logistic regression, linear discriminant analysis, etc., which are difficult to accurately depict the nonlinear relationship between complex ADC-enhancement patterns;

[0007] There is a lack of unified, stable and medically sensitive discrimination function, which cannot take into account the coupling relationship between image features and classification robustness, and is prone to misjudgment;

[0008] There is no complete automatic diagnosis process, which cannot be directly embedded in existing imaging systems for real-time auxiliary diagnosis by radiologists.

[0009] Therefore, there is an urgent need for a new automatic identification method that can realize automatic ROI segmentation, fuse key parameters (such as ADC values and enhancement modes), and construct a discriminant function with clear physical meaning and strong discriminant ability, so as to improve the diagnostic accuracy and efficiency of nasal cavity and sinus melanoma and lymphoma in magnetic resonance images. SUMMARY

[0010] The present application aims to solve the problems of strong subjectivity of manual ROI delineation, low utilization efficiency of image features, and low classification accuracy in existing nasal cavity and sinus melanoma and lymphoma magnetic resonance image recognition, and proposes a brain tumor recognition and analysis system and method for magnetic resonance images, which realizes efficient and accurate differentiation of the two diseases through automatic ROI segmentation and a nonlinear discriminant function with clear physical meaning and strong robustness, and provides a new solution for clinical auxiliary diagnosis.

[0011] To achieve the above-mentioned purpose, the present application provides the following technical solutions: a brain tumor recognition and analysis method for magnetic resonance images, characterized in that it comprises the following steps:

[0012] S1. Obtain DCE-MRI and DWI sequence image data of the subject;

[0013] S2. Preprocess the obtained image data, including noise suppression, registration, and standardization; S3. Automatically extract the ROI region of the tumor based on an image segmentation model;

[0014] S4. Extract the ADC value from the ROI region of the DWI image and the enhancement mode from the ROI region of the DCE-MRI image;

[0015] S5. Based on the extracted ADC value and enhancement mode, construct a double-well potential function as a discriminant model, and calculate the energy difference ;

[0016] S6. If , it is judged to be melanoma, otherwise it is lymphoma, wherein is the energy threshold.

[0017] Preferably, the extraction of the ROI region uses a deep convolutional neural network based on the U-Net structure, the input is a multi-modal MRI image, and the output is a binary tumor mask.

[0018] Preferably, the enhancement mode is automatically classified as type I (inflow type), type II (outflow type), or type III (plateau type) according to the shape of the time-signal intensity curve (TIC), and is corresponded to the enhancement index F: type I F=1, type II F=2, and type III F=3.

[0019] Preferably, the ADC value is obtained in the following manner:

[0020] ADC images are constructed under the conditions of b=0 and b=1000s / mm², the ROI region of the maximum solid tumor component is taken, the cystic or necrotic area is avoided, and the average ADC value is calculated.

[0021] Preferably, the semi-quantitative parameters corresponding to the enhancement mode include: peak time, peak intensity, initial intensity, and change trend after reaching the peak.

[0022] Preferably, the energy threshold Determined through ROC curve analysis.

[0023] Preferably, the image analysis process can be embedded in a PACS system, and has image import, automatic analysis, and result visualization output functions, and supports doctor interaction correction.

[0024] A system adopting any one of the above brain tumor recognition analysis methods of magnetic resonance images, characterized in that the system comprises:

[0025] An image acquisition module is configured to acquire DCE-MRI and DWI sequence image data of a subject;

[0026] An image preprocessing module is configured to preprocess the acquired image data, including noise suppression, registration, and standardization;

[0027] An ROI automatic segmentation module is configured to automatically extract the ROI region of the tumor area based on an image segmentation model;

[0028] A feature extraction module is configured to extract an ADC value from the ROI region of the DWI image and an enhancement mode from the ROI region of the DCE-MRI image, respectively;

[0029] A classification and discrimination module is configured to construct a double-well potential function as a discrimination model based on the extracted ADC value and enhancement mode, and calculate an energy difference , and perform automatic classification of melanoma or lymphoma based on an energy threshold .

[0030] An interaction and output module is configured to visualize the classification result and support connection with a PACS system, and provide a doctor interaction modification interface.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] An automatic ROI extraction process is realized, manual intervention is eliminated, and image labeling efficiency and consistency are improved;

[0033] A non-linear discrimination function based on a double-well potential model is innovatively constructed, which can distinguish complex distribution of parameter features without training a complex model.

[0034] Coupling the ADC value with the DCE enhancement mode modeling is closer to the actual tumor physiological process, and improves the discrimination robustness.

[0035] The whole process has good embeddability, can be integrated into a PACS system, and provides an accurate auxiliary diagnostic tool for an image department doctor.

[0036] Through preliminary test verification, the classification accuracy is better than that of the existing linear model, and has clinical promotion potential.

[0037] In summary, the application provides a brain tumor recognition analysis system and method of a magnetic resonance image, provides an accurate auxiliary diagnostic tool for an image department doctor, and has a wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of a brain tumor recognition analysis method of a magnetic resonance image;

[0039] Figure 2 is a ROI region and signal-time curve of a tumor region;

[0040] Figure 3 is a framework diagram of a brain tumor recognition analysis system of a magnetic resonance image. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all.

[0042] Referring to the accompanying Figure 1 The application relates to artificial intelligence image recognition, more particularly to the field of medical image recognition and analysis, and particularly relates to a brain tumor recognition analysis method of a magnetic resonance image.

[0043] S1. Obtain DCE-MRI and DWI sequence image data of a subject.

[0044] Specifically, the purpose of this step is to provide necessary multi-modal MRI data input for subsequent discriminant analysis, specifically including:

[0045] 1. Subject selection: select a subject from a patient suspected to have a nasal cavity and nasal sinus malignant tumor, and collect image data of the subject.

[0046] 2. MRI scanning protocol:

[0047] DWI sequence (diffusion weighted imaging): a single-shot SE-EPI sequence is used, and common parameter settings are as follows:

[0048] b-value: 0 and 1000 s / mm²; TR / TE: 3000-5000 ms / 60-90 ms; slice thickness: 3~5 mm, no gap; coverage: from nasal cavity to maxillary and ethmoid sinuses.

[0049] DCE-MRI sequence (dynamic contrast-enhanced): using three-dimensional fast gradient echo sequence (such as 3D-FSPGR or VIBE), injecting contrast agent (Gd-DTPA, dose 0.1 mmol / kg, rate 2.5 mL / s), collecting dynamic images at about 5~6 time points, common parameter settings are:

[0050] Baseline + multi-phase arterial, venous and delayed phase images; time interval: 10~15 seconds; resolution: about 1×1×2 mm.

[0051] 3. Data format and export: after acquisition, all DWI and DCE-MRI images are exported in DICOM format for subsequent image processing module.

[0052] S2. Preprocess the acquired image data, including noise suppression, registration and standardization.

[0053] Specifically, to ensure the accuracy of subsequent image segmentation, parameter extraction and discriminant analysis, the following preprocessing operations are required for the acquired images:

[0054] 1. Noise suppression (Denoising):

[0055] Use non-local mean filter (NLM) or bilateral filter for DWI images to preserve edge information;

[0056] Use time series denoising method (such as low-rank SVD filter) for DCE sequence to process dynamic enhancement images and reduce random noise caused by motion or low signal.

[0057] 2. Image registration (Registration):

[0058] Perform rigid or affine registration of DCE images with baseline images;

[0059] Perform multi-modal mutual information registration of DCE images with DWI images to unify the spatial coordinate system and ensure consistency of ROIs between multiple sequences;

[0060] Registration tools can be selected from mature platforms such as ITK, ANTs or Slicer.

[0061] 3. Standardization processing:

[0062] Intensity normalization: Intensity normalization (e.g. Z-score or Min-Max standardization) is applied to ADC images to unify the image contrast across different scanners and subjects.

[0063] Spatial normalization: Images are resampled to a uniform resolution (e.g. 1 mm3 voxels) for input into deep learning models.

[0064] Crop and reposition: Automatic cropping of the head region is performed to remove irrelevant background areas, accelerating processing speed and reducing error propagation.

[0065] After the above preprocessing, the image data has good spatial and grayscale consistency, which can be directly used for automatic ROI segmentation and parameter extraction analysis.

[0066] S3. Automatically extract the ROI region of the tumor based on the image segmentation model.

[0067] Specifically, this step aims to use artificial intelligence methods to automatically segment the tumor region from the multi-modal MRI images of the subject, to obtain the precise region of interest (ROI) required for subsequent analysis.

[0068] Preferably, the extraction of the ROI region uses a deep convolutional neural network based on the U-Net structure, with the input being multi-modal MRI images and the output being a binary tumor mask. The specific implementation process is as follows:

[0069] 1. Network structure design:

[0070] The deep convolutional neural network with an improved U-Net (Unet) structure is used as the main architecture of the segmentation model, with an encoder-decoder structure and adding skip connections to realize multi-layer feature fusion. The input and output forms of the network are as follows:

[0071] Input: Multi-modal MRI images, including b=1000 images of DWI, ADC images, and multiple time point images (such as early enhancement images, peak enhancement images, delay images, etc.) in DCE-MRI, which are input as input tensors through channel concatenation. The input size is uniformly resampled to (256x256xN), where N is the number of modalities or time points.

[0072] Output: Binary tumor mask image (binary mask) with the same size as the input, where the region with a value of 1 represents the predicted tumor region, and the region with a value of 0 is the background.

[0073] 2. Model training and optimization:

[0074] Training data: Using the labeled paranasal sinus tumor image dataset, the doctors manually delineate the tumor boundary as the ground truth.

[0075] Loss function: The weighted combination of Dice loss function and cross-entropy loss function can improve the sensitivity of the model to small tumor regions:

[0076]

[0077] where, is the total loss function used for network training; is the Dice loss function, which measures the overlap between the predicted region and the true region; is the cross-entropy loss function, which is used for pixel-level classification error measurement; and is the weighted system used to balance the contribution of the two loss terms. Usually, , the commonly used setting is .

[0078] The Dice loss function and the cross-entropy loss function use the conventional settings in the field, and this application does not make further introduction.

[0079] Data augmentation: In the training phase, image rotation, scaling, mirroring, pseudo-random cropping, contrast disturbance, etc. Data augmentation strategies are used to enhance the generalization ability of the model.

[0080] Optimizer and learning rate strategy: Use Adam optimizer, initial learning rate set to 0.001, use Cosine Annealing method for adaptive decay during training.

[0081] 3. Inference and post-processing:

[0082] Prediction: After the model training is completed, forward inference (Inference) is performed on the new input MRI image to generate the tumor mask.

[0083] Post-processing: Morphological processing (such as closing operation, connected component analysis) is performed on the predicted mask to remove isolated false positive regions; select the connected region with the largest area or located in the nasal cavity / nasal sinus anatomical region as the final ROI, ROI region refer to the 2E, 3E frame line part in the attached Figure 2 .

[0084] 4. Multi-modal fusion strategy (optional enhancement):

[0085] In the encoder part, shared weights or feature attention mechanisms (SE block / CBAM) are used to fuse different modal features and improve segmentation accuracy. Multi-scale feature fusion modules (such as ASPP, deep separable convolution) can be introduced to improve the ability to depict tumor boundaries.

[0086] S4. Extract the ADC value from the ROI region of the DWI image, and extract the enhancement mode from the ROI region of the DCE-MRI image.

[0087] The specific steps are as follows:

[0088] 1. Extract the ADC value:

[0089] From the DWI image, use the tumor ROI mask obtained by automatic segmentation to complete the following processing:

[0090] Image source: Use b=0 s / mm² and b=1000 s / mm² DWI images to construct ADC maps (Apparent Diffusion Coefficient Map) by system calculation.

[0091] Calculation method: According to the intensity of the DWI image Exponential decay formula with b value change:

[0092]

[0093] Where, is the baseline signal intensity of the point at b=0; ADC is the apparent diffusion coefficient, which is an index that quantitatively reflects the diffusion ability of water molecules in the tissue. The larger it is, the more free the diffusion is; the smaller it is, the more limited the diffusion is.

[0094] The system performs fitting calculation on each pixel point to generate an ADC image.

[0095] ROI extraction: In the ADC image, use the ROI mask obtained in S3 to extract the ADC values of all pixels in the tumor region.

[0096] Result representation: The ADC values in the ROI are statistically calculated to calculate their average value, denoted as A. In this invention, the ADC value A will be one of the key variables in the discriminant function, with the unit of .

[0097] 2. Extract the enhancement mode:

[0098] Extract the time-signal intensity curve (Time-Intensity Curve, TIC) reflecting the perfusion characteristics of the tumor from the DCE-MRI image sequence, and automatically classify it.

[0099] Preferably, the enhancement pattern is automatically classified as Type I (inflow), Type II (outflow) or Type III (plateau) according to the shape of the time-intensity curve (TIC), and is corresponded to the enhancement index F: Type I F=1, Type II F=2, Type III F=3.

[0100] Preferably, the semi-quantitative parameters corresponding to the enhancement pattern include: time to peak, peak intensity, initial intensity, and the trend after reaching the peak.

[0101] Image processing procedure: using the tumor ROI obtained in S3, the average signal intensity of the ROI region is extracted on each time-phase image of DCE-MRI; according to the acquisition time sequence, the signal-time curve of the ROI region, i.e. the TIC curve, is constructed.

[0102] Curve feature analysis: the following indicators of the curve are analyzed by the system:

[0103] Time to peak ; Peak intensity ; Initial intensity ; Trend after reaching the peak (rising, falling or plateau).

[0104] Enhancement pattern classification: the system automatically classifies according to the shape of the TIC curve, and the rules are as follows:

[0105]

[0106] The classification adopts rule-driven combined with curve slope judgment:

[0107] Let the early slope be , where is the peak signal intensity, is the initial intensity, is the time point corresponding to , i.e. the time to peak, and correspond to the time points; it reflects the rate at which the signal increases from to after the contrast agent enters the tumor blood vessels, reflecting the tumor blood perfusion speed and capillary permeability;

[0108] The late slope is , where is the delayed signal intensity, and correspond to the time points; it depicts the rate at which the signal decreases (or plateaus) with time from the peak to the delay phase, reflecting the speed of contrast agent washing out of the tissue and the characteristics of liquid exchange inside and outside the capillary;

[0109] The judgment criteria are:

[0110] If : Type III;

[0111] If : Type II;

[0112] Otherwise, Type I. Referring to the curves of 2F in FIG. 2 and 3F in FIG. 3, the curve of 2F is of Type II and the curve of 3F is of Type III. Figure 2 The reinforcement index is assigned: the classification result is corresponded to a reinforcement index

[0113] , which is used for subsequent discriminant function construction. The reinforcement index is a dimensionless discrete integer variable (1-3).

[0114] S5. Based on the extracted ADC value and the reinforcement mode, a double-well potential function is constructed as a discriminant model, and the energy difference is calculated.

[0115] Specifically, the core idea of this step is derived from the double-well potential model in physics, which is often used to describe the transition process of a system between two stable states. Inspired by this, the present application proposes to model the nasal cavity and sinus melanoma and lymphoma as two "energy potential wells", respectively, and to construct an energy function using the ADC value and the reinforcement mode index F as input variables to achieve discriminant classification.

[0116] Preferably, the double-well potential function is defined as:

[0117]

[0118] wherein: A is the normalized ADC value, F is the reinforcement index, : are the typical ADC value centers of representative lymphoma and melanoma, respectively; : are the typical reinforcement index centers of representative lymphoma and melanoma, respectively; : is an adjustment coefficient.

[0119] Preferably, the ADC value is obtained in the following manner:

[0120] Under the conditions of b=0 and b=1000s / mm², the ADC image is constructed, the ROI region of the largest solid tumor component is taken, the cystic or necrotic area is avoided, and the average ADC value is calculated.

[0121] 1. Constructing a double-well potential function

[0122] Let the input variables be:

[0123] : the average ADC value of the ROI region, in ;​

[0124] : the index value corresponding to the enhancement mode, taking values in {1, 2, 3}.

[0125] The double-well potential function is introduced:

[0126]

[0127] where:

[0128] : the typical ADC value centers of representative lymphoma and melanoma, respectively;

[0129] : the typical enhancement index centers of representative lymphoma and melanoma, respectively;

[0130] : the adjustment coefficient, used to adjust the contribution ratio of ADC and enhancement mode to the energy function.

[0131] Specifically, the adjustment coefficient can be set based on the variance.

[0132] First, the overall variance of A and F is calculated on the entire training set respectively , and the adjustment coefficient is inversely proportional to the variance:

[0133] , . In this way, the dimension and numerical range of the two terms in the energy function can be balanced.

[0134] This function has two "potential well" centers, located at and , and the lowest points can be regarded as the "energy ground state" of the two pathological types.

[0135] 2. Calculate the energy difference ΔE

[0136] The energy difference ΔE is calculated as follows:

[0137]

[0138] where is the symmetry center of the double-well function, and if is greater than the preset threshold , it is determined as the corresponding disease type.

[0139] S6. If , it is determined as melanoma, otherwise as lymphoma, where is the energy threshold.

[0140] Specifically, an energy threshold is set ;

[0141] The discrimination rule is as follows:

[0142] If : it indicates that the test point is more inclined to melanoma energy sink in the energy space → it is discriminated as melanoma;

[0143] Otherwise (i.e. ): it is discriminated as lymphoma.

[0144] This way has the following technical effects:

[0145] Strong non-linear interpretability: the double-well energy function does not depend on simple linear weighting, but abstractly models the "energy state" between the two lesions through the potential well structure;

[0146] High robustness: the use of square terms and product terms enhances the modeling ability of the joint influence of ADC values and enhancement methods;

[0147] Suitable for small sample learning: the structure does not need a large number of parameter training, and is convenient for deployment in medical images.

[0148] Preferably, the energy threshold is determined by ROC curve analysis.

[0149] Specifically, the energy threshold is determined based on ROC curve analysis, and the specific process is as follows:

[0150] 1. Sample and energy difference data preparation

[0151] A group of diagnosed nasal sinus tumor cases (melanoma and lymphoma) are collected, a total of cases.

[0152] For each case, calculate its energy difference according to the foregoing steps.

[0153] At the same time, record the true label of each case :

[0154]

[0155] 2. Draw the ROC curve

[0156] All are taken as a group of predicted scores and true labels.

[0157] Scan the possible threshold (from the minimum to the maximum ):

[0158] For each The discriminant rule is set as:

[0159]

[0160] The corresponding sensitivity (True Positive Rate, TPR) and specificity (Specificity, SP) are calculated, wherein

[0161]

[0162]

[0163] The are sequentially connected to obtain the ROC curve.

[0164] 3. Select the optimal threshold

[0165] The threshold corresponding to the Youden index is calculated:

[0166]

[0167] Find the threshold that makes the maximum:

[0168]

[0169] The at this time is the energy difference discriminant threshold, and the comprehensive performance of sensitivity and specificity is best at this point.

[0170] 4. Apply discriminant

[0171] Calculate for new samples,

[0172] If , it is determined as melanoma;

[0173] If , it is determined as lymphoma.

[0174] The technical effects of determining the energy threshold based on the ROC curve analysis are as follows:

[0175] Determine through large sample ROC analysis, without subjective setting, taking into account sensitivity and specificity; redo the same process in different data sets or centers to obtain the optimal threshold for the population.

[0176] ​Preferably, the image analysis process can be embedded in a PACS system, with image import, automatic analysis, and result visualization output functions, supporting physician interactive correction.

[0177] In particular, the image analysis process can be embedded in a PACS system (Picture Archiving and Communication System), providing an integrated image management and analysis platform. Specifically, this function enhances the ability of image import, automatic analysis, result output, and interactive correction, etc., to improve the efficiency and accuracy of medical image diagnosis.

[0178] 1. Image import function:

[0179] This method can be seamlessly integrated with existing PACS systems, realizing automatic import and management of MRI images of different sources (including DCE-MRI and DWI images). The system can receive image data collected from different devices (such as MRI scanners from multiple hospitals and different manufacturers), and automatically identify the metadata of the images (such as sequence type, resolution, scanning parameters, etc.).

[0180] 2. Automatic analysis function:

[0181] The system can automatically call the image analysis algorithm process described in the present application to process the imported MRI images. Automatic ROI segmentation: Based on deep learning models such as U-Net, automatically extract the tumor region (ROI). ADC value and enhancement mode extraction: Automatically extract ADC values and enhancement types from DWI and DCE-MRI images, and calculate the energy difference ΔE based on the double-well potential function.

[0182] Automatic classification and discrimination: According to the energy difference ΔE and the threshold , automatically determine whether the tumor is melanoma or lymphoma.

[0183] 3. Result visualization output function:

[0184] Visualization results: After analysis, the system can present multiple levels of analysis results in the interface: Image superposition: Show the superposition results of the original MRI image and the automatically extracted ROI region, which is convenient for doctors to directly observe the tumor region. ADC and enhancement index map: Display the visualization results of ADC map and TIC curve, enhancement type, and intuitively present the biological characteristics of the tumor. Discrimination results: Output the tumor type classification results based on ΔE calculation, as well as the related probability or confidence index. Image / data report: Generate a comprehensive report including diagnosis results, image data, statistical analysis, etc., supporting export to standardized formats (such as PDF, DICOM, etc.).

[0185] 4. Support doctor interaction correction function:

[0186] Interactive interface: the system provides an interactive interface for doctors to correct and confirm the automatic analysis results. For example, doctors can manually adjust the ROI region to optimize the recognition of tumor boundaries. Doctors can modify or verify the extracted ADC values and enhancement patterns, and manually review the classification results. Feedback mechanism: the system can learn from the doctor's corrections and feedback, gradually optimizing the algorithm's performance in subsequent image analysis, thereby improving the accuracy of automatic analysis.

[0187] Referring to the accompanying drawings Figure 3 A system using any of the above brain tumor recognition and analysis methods of magnetic resonance images, characterized in that the system comprises:

[0188] An image acquisition module for acquiring DCE-MRI and DWI sequence image data of a subject;

[0189] An image preprocessing module for preprocessing the acquired image data, including noise suppression, registration and standardization;

[0190] An ROI automatic segmentation module for automatically extracting the ROI region of the tumor region based on an image segmentation model;

[0191] A feature extraction module for extracting ADC values from the ROI region of the DWI image and extracting enhancement patterns from the ROI region of the DCE-MRI image, respectively;

[0192] A classification and discrimination module for constructing a double-well potential function as a discrimination model based on the extracted ADC values and enhancement patterns, and calculating the energy difference And based on the energy threshold Automatically classify melanoma or lymphoma;

[0193] An interaction and output module for visualizing and outputting the classification results, and supporting interface with the PACS system to provide a doctor interaction modification interface.

[0194] Compared with the prior art, the present application has the following beneficial effects:

[0195] The automatic ROI extraction process is realized, manual intervention is eliminated, and the image annotation efficiency and consistency are improved;

[0196] Innovatively, a non-linear discrimination function based on a double-well potential model is constructed, which can distinguish complex distribution of parameter features without training complex models;

[0197] Coupling modeling of ADC values and DCE enhancement patterns is closer to the actual tumor physiological process, and improves the discrimination robustness;

[0198] The whole process has good embeddability and can be integrated into a PACS system to provide an accurate auxiliary diagnosis tool for an image doctor.

[0199] Through preliminary test verification, the classification accuracy is better than that of an existing linear model, and has clinical promotion potential.

[0200] In conclusion, the application provides a brain tumor recognition and analysis system and method for magnetic resonance images, which provides an accurate auxiliary diagnosis tool for an image doctor and has wide application prospects.

[0201] The above merely describes a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and inventive concept of the application within the technical range disclosed by the application, which should be covered in the protection scope of the application.

Claims

1. A method for identifying and analyzing brain tumors from magnetic resonance images, characterized in that, Includes the following steps: S1. Acquire DCE-MRI and DWI sequence image data of the subject; S2. Preprocess the acquired image data, including noise suppression, registration, and normalization; S3. Automatically extract the ROI region of the tumor based on the image segmentation model; S4. Extract ADC values ​​from the ROI regions of DWI images and extract enhancement patterns from the ROI regions of DCE-MRI images; S5. Based on the extracted ADC values ​​and enhancement methods, a dual-well potential energy function is constructed as a discrimination model, and the energy difference is calculated. ; The dual-well potential energy function Defined as: Where: A is the standardized ADC value, and F is the enhancement index. These are the typical ADC value centers for representative lymphomas and melanomas, respectively; These are the typical enhancement index centers for representative lymphomas and melanomas, respectively; : is the adjustment coefficient; The formula for calculating the energy difference ΔE is: in Let be the center of symmetry of the double-well function, if Greater than the preset threshold If so, it is identified as the corresponding disease type; S6. If If it is positive, it is diagnosed as melanoma; otherwise, it is lymphoma. This is the energy threshold.

2. The method according to claim 1, characterized in that, The ROI region extraction employs a deep convolutional neural network based on a U-Net structure, with multimodal MRI images as input and a binary tumor mask as output.

3. The method according to claim 1, characterized in that, The enhancement method is automatically classified into Type I inflow, Type II outflow, or Type III platform based on the shape of the Time-Signal Intensity (TIC) curve, and corresponding to the enhancement index F: Type I is F=1, Type II is F=2, and Type III is F=3.

4. The method according to claim 1, characterized in that, The ADC value is obtained as follows: ADC images were constructed under the conditions of b=0 and b=1000s / mm², the ROI region of the layer where the largest solid tumor component is located was selected, avoiding cystic or necrotic areas, and the average ADC value was calculated.

5. The method according to claim 1, characterized in that, The semi-quantitative parameters corresponding to the enhancement method include: peak time, peak intensity, initial intensity, and the trend of change after peak.

6. The method according to claim 1, characterized in that, Energy threshold Determined through ROC curve analysis.

7. The method according to any one of claims 1 to 6, characterized in that, The image analysis workflow can be embedded in the PACS system, featuring image import, automatic analysis, and result visualization output, and supports interactive correction by doctors.

8. A system employing the brain tumor identification and analysis method based on magnetic resonance images as described in any one of claims 1 to 7, characterized in that, The system includes: The image acquisition module is used to acquire DCE-MRI and DWI sequence image data of the subject; The image preprocessing module is used to preprocess the acquired image data, including noise suppression, registration, and normalization. The ROI automatic segmentation module is used to automatically extract the ROI region of the tumor area based on the image segmentation model. The feature extraction module is used to extract ADC values ​​from the ROI region of DWI images and the enhancement mode from the ROI region of DCE-MRI images, respectively. The classification and discrimination module is used to construct a dual-well potential energy function as a discrimination model based on the extracted ADC values ​​and enhancement methods, and to calculate the energy difference. And based on energy threshold Automatic classification of melanoma or lymphoma; The interaction and output module is used to visualize the classification results and supports integration with PACS systems, providing doctors with an interactive modification interface.

Images