Lung blood flow perfusion automatic classification method, device and program product based on dual-energy CT (Computed Tomography)

Abstract

The invention relates to the field of intelligent medical treatment, in particular to a lung blood flow perfusion automatic classification method and device based on dual-energy CT and a program product. Comprising the steps that lung dual-energy CT scanning data of a to-be-tested person is acquired, and the dual-energy CT scanning data comprises a CT image and a PBV image; performing lung segmentation on the CT image to obtain a segmented CT image; carrying out registration on the segmented CT image and the PBV image to obtain a registered functional image; performing function extraction based on the registered function image to obtain function information; and performing lung function region division through the function information to obtain a division result. By combining the dual-energy CT image and the deep learning technology, a rapid, accurate and non-invasive lung perfusion evaluation tool is provided for clinicians, and the method is particularly suitable for diagnosis and identification of pulmonary vascular diseases and has good clinical value.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical treatment, and specifically to a method, device, program product and computer-readable storage medium for automatic classification of pulmonary blood perfusion based on dual-energy CT. Background Art

[0002] Pulmonary blood perfusion is one of the key functions of the lungs, and blood perfusion assessment plays a vital role in the diagnosis, treatment, and prognosis of respiratory diseases. Traditional pulmonary blood perfusion assessment methods, such as lung perfusion testing, can provide important physiological information, but often fail to provide detailed spatial distribution information of the internal structure and function of the lungs. In addition, some patients may not be able to complete routine lung perfusion testing, or the test results may be affected by a variety of factors.

[0003] In the diagnosis of pulmonary vascular diseases, the blood perfusion of acute pulmonary embolism and chronic thromboembolic pulmonary hypertension is closely related to the patient's prognosis. The secondary pulmonary blood perfusion of acute pulmonary embolism is highly correlated with the clinical risk of acute pulmonary embolism. The lower the vascular perfusion, the higher the risk of acute right heart failure and short-term death in patients. The corresponding quantitative evaluation of pulmonary thrombus load and pulmonary blood perfusion will be beneficial to the clinical risk assessment and treatment strategy selection of acute pulmonary embolism. In patients with chronic thromboembolic pulmonary hypertension, pulmonary emboli and vascular perfusion are not matched, and blood perfusion is closely related to pulmonary vascular remodeling, increased pulmonary vascular resistance, and decreased right heart function. The assessment of the degree of pulmonary blood perfusion impairment can be used for non-invasive assessment of hemodynamics and right heart function. Therefore, pulmonary vascular perfusion in patients with chronic thromboembolic pulmonary hypertension is an important supplement to hemodynamic assessment. Chronic obstructive pulmonary disease (COPD) and interstitial lung disease (ILD) cause pulmonary vascular damage and abnormal blood perfusion, which are closely related to the severity of the disease and the patient's prognosis. Quantitative analysis of chronic pulmonary blood perfusion is of great value in guiding treatment and monitoring the disease. However, current blood perfusion assessment mainly uses radionuclide pulmonary ventilation perfusion imaging, which requires two radionuclide imagings and has low spatial resolution and poor quantitative analysis performance. It also cannot accurately display diseased blood vessels. Summary of the invention

[0004] In recent years, the development of medical imaging technology has provided new possibilities for pulmonary blood perfusion assessment. In particular, the emergence of dual-energy CT technology enables the acquisition of high-resolution anatomical information while also providing functional information. In view of the above problems, the present invention proposes an automatic classification method for pulmonary blood perfusion based on dual-energy CT, which specifically includes:

[0005] Acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images;

[0006] Performing lung segmentation on the CT image to obtain a segmented CT image;

[0007] Registering the segmented CT image with the PBV image to obtain a registered functional image;

[0008] Performing function extraction based on the registered functional image to obtain functional information;

[0009] The lung functional area is divided according to the functional information to obtain a division result.

[0010] The regional division includes perfusion defect area, perfusion reduction area, and healthy area;

[0011] Optionally, the division is to input the functional information and the functional image into a classifier for division to obtain a division result;

[0012] Optionally, the functional information includes local blood flow, blood volume, and mean transit time;

[0013] Optionally, the process of dividing by a classifier is:

[0014] Obtain functional images and functional information;

[0015] Performing feature conversion on the functional information to obtain functional features;

[0016] Extract spatial distribution features based on functional images;

[0017] Input the functional features and spatial distribution features into the classifier for classification to obtain the classification results;

[0018] Optionally, the process of dividing by the classifier further includes region elimination, and after the region elimination, a region result after removal is obtained, and evaluation is performed based on the region result after removal; the region elimination is to perform morphological processing on the region in the division result to remove isolated small regions to obtain the region result after removal;

[0019] Optionally, the classifier includes one or more of the following: random forest, decision tree, support vector machine, extreme learning machine, perceptron vector machine, convolutional neural network, residual network, and hole convolutional neural network.

[0020] The division is obtained by comparing the local blood flow with a preset threshold value;

[0021] Optionally, the preset threshold of local blood flow in the perfusion defect area is smaller than that in the perfusion reduced area, and the preset threshold of local blood flow in the perfusion reduced area is smaller than that in the healthy area;

[0022] Optionally, when the local blood flow is less than 10 ml / 100 ml / min, it is determined as a perfusion defect area, and when the local blood flow is greater than or equal to 30 ml / 100 ml / min, it is determined as a healthy area;

[0023] Optionally, the division is first performed by comparing the local blood flow with a preset threshold to obtain a divided area, and then the divided area and functional information are input into a classifier to perform fine regional division to obtain a division result.

[0024] The registration is performed by combining a rigid and non-rigid algorithm to register the segmented CT image and the PBV image; a rigid algorithm is first used to perform a rough registration to obtain a rough registration result, and then a non-rigid algorithm is used to perform local shape constraints on the rough registration result to obtain a registered functional image;

[0025] Optionally, the registration further includes fine adjustment, and after performing local shape constraints by a non-rigid algorithm, the local area is precisely aligned by a block matching algorithm to obtain a registered functional image;

[0026] Optionally, the rigid algorithm calculates rigid transformation parameters between the segmented CT image and the PBV image and performs rough registration based on the rigid transformation parameters to obtain a rough registration result;

[0027] Optionally, the rigid algorithm further includes parameter optimization, optimizing the rigid transformation parameters by gradient descent to obtain optimized rigid transformation parameters, and performing rough registration based on the optimized rigid transformation parameters to obtain a rough registration result;

[0028] Optionally, the non-rigid algorithm performs local deformation constraints through a B-spline deformation model;

[0029] Optionally, the finely adjusted regions include: lung lobe boundaries, vascular structures;

[0030] Optionally, the registered functional image is quantitatively evaluated through a structural similarity index and a registered mutual information value. When the quantitative evaluation result is unqualified, re-registration or manual adjustment is performed until the quantitative evaluation result is qualified.

[0031] The segmentation is performed by a pre-trained segmentation model to obtain a segmentation result, and the segmentation result includes a left lung and a right lung;

[0032] Optionally, the segmentation model includes one or more of the following: U-Net, FCN, SegNet, DeepLabV3+, ENet;

[0033] Optionally, the segmentation model includes an encoder and a decoder, the encoder consists of N residual blocks, N is a natural number greater than 1, the input data is extracted through the N residual blocks connected in series in the encoder to obtain multi-scale features, the features of the Nth residual block are input to the decoder, and the first N-1 multi-scale features are input to the decoder through jump connections to fuse features of different scales to obtain decoding features, and then the decoding features are weighted by the attention mechanism to obtain important areas, and the important areas are segmented to obtain segmentation results.

[0034] The method further includes data preprocessing, performing data preprocessing on the CT image to obtain a processed CT image, and segmenting the processed image to obtain a segmented CT image;

[0035] Optionally, the data preprocessing includes pixel normalization, size unification, image noise reduction, content highlighting, and data enhancement;

[0036] Optionally, the content highlights lung tissue by window width and window level;

[0037] Optionally, the image denoising removes image noise by Gaussian filtering;

[0038] Optionally, the segmentation further includes data post-processing, performing data post-processing on the segmented CT image to obtain a post-processed CT image, and registering the post-processed CT image with the PBV image; the post-processing is performing morphological processing on the segmented CT image to obtain the post-processed CT image;

[0039] Optionally, the morphological processing includes opening and closing operations, removing isolated areas, and filling holes;

[0040] Optionally, the post-processing further includes edge optimization, performing edge optimization on the morphologically processed CT image to obtain a post-processed CT image;

[0041] Optionally, the edge optimization is performed by conditional random fields.

[0042] The object of the present invention is to provide a method for evaluating pulmonary blood perfusion based on dual-energy CT, comprising:

[0043] Acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images;

[0044] The segmentation results were obtained based on the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT;

[0045] A pulmonary blood perfusion assessment is performed based on the segmentation result to obtain an assessment result.

[0046] The object of the present invention is to provide a computer product, including a computer program or instructions, which are executed by a processor to implement the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT or to execute the above-mentioned pulmonary blood perfusion evaluation method based on dual-energy CT.

[0047] The object of the present invention is to provide a computer device, comprising a memory, a processor and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT or to execute the above-mentioned pulmonary blood perfusion evaluation method based on dual-energy CT.

[0048] The object of the present invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT or to execute the above-mentioned pulmonary blood perfusion evaluation method based on dual-energy CT.

[0049] Advantages of the present invention:

[0050] 1. Innovative diagnostic methods and processes. The present invention proposes a lung perfusion assessment method based on dual-energy CT and deep learning models, which provides a new, rapid and non-invasive diagnostic method for the diagnosis of lung diseases in clinical practice. This method specifically targets the problems of "same disease with different symptoms" and "different diseases with same symptoms" in pulmonary vascular diseases, and adopts an innovative step-by-step diagnostic process. First, the CT image is segmented by a deep learning model to identify the lung area. Then, the segmented CT image is accurately aligned with the PBV image, combining structural and functional information. Finally, through pixel-level analysis of the PBV image, perfusion defects, reduced perfusion and healthy areas are identified. This gradually refined diagnostic process not only improves the accuracy of diagnosis, but also effectively distinguishes easily confused disease types, greatly improving the efficiency of clinical diagnosis.

[0051] 2. Comprehensive use of structural and functional information. The present invention makes full use of the advantages of dual-energy CT technology, and simultaneously obtains CT images reflecting the lung structure and PBV images reflecting the lung function. Through deep learning models and image registration technology, accurate fusion of structural information and functional information is achieved. This comprehensive analysis method can provide more comprehensive and accurate lung perfusion assessment results. For example, CT images can clearly display the anatomical structure of the lungs, including airways, blood vessels, and lung parenchyma, while PBV images can reflect the dynamic changes in pulmonary blood perfusion. The combination of the two types of information enables doctors to simultaneously evaluate the morphological changes and functional status of the lungs, providing a solid foundation for accurate diagnosis and individualized treatment.

[0052] 3. Advanced image processing and artificial intelligence technologies. The present invention adopts a number of advanced image processing and artificial intelligence technologies to improve the accuracy and efficiency of diagnosis. First, a deep learning model (such as an improved pre-trained U-Net network) is used to perform high-precision lung segmentation on CT images, providing an accurate anatomical basis for subsequent analysis. Secondly, a non-rigid registration algorithm (such as a B-spline deformation model combined with a mutual information similarity metric) is used to achieve accurate registration of CT images and PBV images, ensuring the spatial consistency of structural and functional information. In addition, the present invention also introduces an edge detection algorithm to extract the imaging features of abnormal perfusion areas, and uses machine learning methods (such as random forests) to intelligently classify functional areas. The combined use of these technologies not only improves the accuracy of diagnosis, but also greatly reduces the time and subjective errors of manual operations.

[0053] 4. Comprehensive feature extraction and analysis. Considering the high similarity of pulmonary vascular disease images, the present invention adopts whole lung imaging and multi-scale feature extraction strategies. Specifically, the whole lung image is segmented into ROIs to obtain multiple sub-regions, and feature extraction is performed on each sub-region separately. This method not only takes into account global information, but also captures local subtle changes. During the feature extraction process, various types of imaging genomics features are used, including first-order statistical features, shape-based features, texture features, etc. Through feature fusion and dimensionality reduction techniques, a high-dimensional and high-precision feature vector is finally obtained. This comprehensive feature analysis strategy greatly improves the model's ability to recognize tiny lesions, and provides strong support for accurately distinguishing different types of interstitial lung diseases. At the same time, this method also lays the foundation for future longitudinal studies and prognostic assessments, and has the potential to discover new imaging biomarkers. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0055] Figure 1 A schematic flow chart of an automatic classification method for pulmonary blood perfusion based on dual-energy CT provided in an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of an automatic classification system for pulmonary blood perfusion based on dual-energy CT provided in an embodiment of the present invention;

[0057] Figure 3 A schematic diagram of an automatic classification device for pulmonary blood perfusion based on dual-energy CT provided in an embodiment of the present invention;

[0058] Figure 4 The UNet deep learning network structure provided by the embodiment of the present invention;

[0059] Figure 5 The data flow of the solution provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0061] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0062] Figure 1 A schematic diagram of an automatic classification method for pulmonary blood perfusion based on dual-energy CT provided in an embodiment of the present invention specifically includes:

[0063] S101: Acquire dual-energy CT scan data of the lungs of the subject, wherein the dual-energy CT scan data includes CT images and PBV images;

[0064] In one embodiment, the method further includes data preprocessing, performing data preprocessing on the CT image to obtain a processed CT image, and segmenting the processed image to obtain a segmented CT image;

[0065] Optionally, the data preprocessing includes pixel normalization, size unification, image noise reduction, content highlighting, and data enhancement.

[0066] In one embodiment, the content highlights lung tissue via a window width and window level.

[0067] Optionally, the image denoising removes image noise by Gaussian filtering;

[0068] Optionally, the segmentation further includes data post-processing, performing data post-processing on the segmented CT image to obtain a post-processed CT image, and registering the post-processed CT image with the PBV image; the post-processing is performing morphological processing on the segmented CT image to obtain the post-processed CT image;

[0069] Optionally, the morphological processing includes opening and closing operations, removing isolated areas, and filling holes;

[0070] Optionally, the post-processing further includes edge optimization, performing edge optimization on the morphologically processed CT image to obtain a post-processed CT image;

[0071] Optionally, the edge optimization is performed by conditional random fields.

[0072] In a specific embodiment, the dual-energy CT (DE-CT) lung enhanced scanning data of the patient is obtained, including chest CT images and pulmonary blood volume (PBV) images. Among them, the CT image provides structural information of the lung, and the PBV image provides information on the lung perfusion function. The patient starts the DE-CT scanning program in a breath-holding state. The scanning range covers the entire lung, from the apex to the base of the lung. The patients are all injected with 350-370 mg iodine / ml iodinated contrast agent through the elbow vein at a rate of 4-5 ml / second (80 ml for patients with a body weight of ≤80 kg; 100 ml for patients with a body weight of >80 kg). For the dual-source dual-energy CT (DS-DE-CT) equipment, the scanning parameters are set as follows: the parameters are 80 / 150 kV with tin filter and automatic exposure control. The reference tube current is 180 mAs, the pitch is 1.1:1, the scanning frame rotation time is 0.28 seconds, and the detector configuration is 192×0.6. Single-source dual-energy CT (SS-DECT) was performed with 80 / 140 kV fast voltage switching, fixed tube current of 280–360 mA, rotation time of 0.5 s, pitch of 1.375:1, and 64 × 0.625 detector configuration. For all DECT, cross-sectional virtual single-energy (40 keV) and material decomposition iodine (MDI) images were generated.

[0073] In a specific embodiment, data acquisition;

[0074] a) Selection of scanning equipment: Data acquisition is the starting point of the entire lung perfusion assessment process, and its specific implementation details are as follows: Select advanced dual-energy CT scanners, such as DS-DE-CT. These devices can simultaneously acquire CT data at two different energy levels.

[0075] b) Contrast injection:

[0076]

[0077] c) Scan parameter settings:

[0078]

[0079]

[0080] d) Image reconstruction method:

[0081] Iterative reconstruction algorithm Such as ADMIRE or ASiR-V Reconstruction thickness 1mm Reconstruction interval 0.8mm Field of View (FOV) 350mm Image size 512×512

[0082] e) Data storage:

[0083] The patient is scanned using a dual-energy CT scanner to obtain CT images and PBV images of the entire lung. CT images mainly provide detailed structural information of the lungs, including anatomical structures such as the lung lobes, trachea, and blood vessels. PBV images provide pulmonary blood perfusion information. The original data is saved in DICOM format, and a patient database is established, including a patient information database, including age, gender, clinical diagnosis, etc.

[0084] In a specific embodiment, data preprocessing: standardizing the input CT image, including pixel value normalization and image size adjustment, using window width and window level technology to highlight lung tissue, and applying Gaussian filtering to reduce noise.

[0085] Data preprocessing:

[0086] Image normalization: linearly map the CT value to the [0,1] interval;

[0087] Window width and window position adjustment: lung window is set to window width 1500HU, window position -600HU;

[0088] Image resizing: resample all images to 512×512×Z, where Z is the number of slices to fit the network input;

[0089] Data augmentation: Randomly apply rotation (±15°), scaling (0.9-1.1 times), translation (±10%), flipping, etc.

[0090] S102: performing lung segmentation on the CT image to obtain a segmented CT image;

[0091] In one embodiment, the segmentation is performed using a pre-trained segmentation model to obtain a segmentation result, and the segmentation result includes a left lung and a right lung.

[0092] Optionally, the segmentation model includes one or more of the following: U-Net, FCN, SegNet, DeepLabV3+, ENet.

[0093] In one embodiment, the segmentation model includes an encoder and a decoder. The encoder is composed of N residual blocks, where N is a natural number greater than 1. The input data is extracted through the N residual blocks connected in series in the encoder to obtain multi-scale features. The features of the Nth residual block are input to the decoder, and the first N-1 multi-scale features are input to the decoder through jump connections to fuse features of different scales to obtain decoding features. The decoding features are then weighted by the attention mechanism to obtain important areas, and the important areas are segmented to obtain segmentation results.

[0094] In a specific embodiment, the acquired CT image is input into a pre-trained deep learning model, and the CT image is segmented to obtain accurate segmentation results of the left lung and the right lung. The present invention adopts an improved U-Net improved network structure as a deep learning model. The network consists of an encoder, a decoder and a jump connection, which can effectively extract multi-scale features and achieve accurate lung segmentation. The network training uses a CT image data set annotated by radiologists in this hospital, and uses data enhancement technology to expand the training samples, including random rotation, scaling, translation and flipping operations. During the training process, the Dice loss function and the Adam optimizer are used. The initial learning rate is set to 0.001, and the step size attenuation strategy is adopted. The learning rate is reduced by 50% every 80 Epochs.

[0095] Network structure design: Adopt an improved U-Net architecture, including an encoder and a decoder. The encoder uses residual blocks to extract multi-scale features, and the decoder fuses feature information at different levels through skip connections. Add an attention mechanism at the end of the network to strengthen the feature expression of key areas.

[0096] Model training: A large number of annotated CT image datasets are used for training. Data augmentation techniques such as random rotation, scaling, and flipping are used to increase the generalization ability of the model. The loss function combines Dice loss and cross entropy loss to balance the class imbalance problem. The Adam optimizer is used, and the learning rate decay strategy is adopted.

[0097] Post-processing optimization: Morphological operations are performed on the segmentation results output by the network, such as opening and closing operations, to remove small isolated areas and fill holes. Conditional random fields (CRF) are used to optimize edges to improve segmentation accuracy. Finally, connected domain analysis is applied to ensure the integrity of the left and right lungs.

[0098] In a specific embodiment, the acquired CT images are input into a pre-trained deep learning model to achieve automatic segmentation of the left lung and the right lung. This embodiment adopts an improved pre-trained U-Net network structure, such as Figure 4 As shown in Figure 2, the network consists of an encoder, a decoder, and skip connections, which can effectively extract multi-scale features and achieve accurate lung segmentation.

[0099] Network structure description:

[0100] Encoder: 4 downsampling blocks, each block contains two 3×3 convolutional layers and one 2×2 max pooling layer;

[0101] Decoder: 4 upsampling blocks, each block contains a 2×2 deconvolution layer and two 3×3 convolution layers;

[0102] Skip connection: concatenate the feature map of the corresponding layer of the encoder with the feature map of the decoder;

[0103] Attention mechanism: Add a spatial attention module to each decoder block to enhance the feature expression of key areas;

[0104] Deep supervision: Add auxiliary losses to feature maps of different scales to accelerate training convergence;

[0105] Activation function: ReLU;

[0106] Final output layer: 1×1 convolutional layer, using Sigmoid activation function;

[0107] Loss function: Combine multiple loss functions to improve segmentation accuracy:

[0108] Dice loss: Ldice=1-(2|X∩Y|+ε) / (|X|+|Y|+ε);

[0109] Cross entropy loss: CE = -Σ[y log(p) + (1-y)log(1-p)];

[0110] Boundary loss:

[0111] Total loss: Ltotal = λ1Ldice + λ2CE + λ3Lb, where λ1, λ2, and λ3 are weight coefficients;

[0112] Training strategy:

[0113] Dataset: A 500-case CT dataset with expert annotations was used, of which 400 were used for training and 100 for validation;

[0114] Optimizer: Adam optimizer, with the initial learning rate set to 0.001;

[0115] Learning rate adjustment: cosine annealing strategy is adopted, and the minimum learning rate is 1e-6;

[0116] Batch size: 32 (adjusted according to GPU memory);

[0117] Training rounds: 200 rounds, early stopping strategy based on the Dice coefficient of the validation set;

[0118] Post-processing:

[0119] Connected domain analysis: remove areas with a volume less than 50 mm 3 isolated areas;

[0120] Morphological operation: Apply a spherical structure element with a radius of 2 voxels to perform an opening operation to remove small protrusions;

[0121] Cavity filling: Filling of small cavities within the lung parenchyma;

[0122] Separation of left and right lungs: Use the watershed algorithm to separate the adhered left and right lungs;

[0123] Model Evaluation:

[0124] Evaluation indicators: Dice coefficient, Hausdorff distance, average surface distance;

[0125] Cross-validation: Perform 5-fold cross-validation to ensure the stability and generalization ability of the model;

[0126] Comparison with manual annotation: The consistency of the segmentation results of the calculation model and the manual segmentation results of two radiologists;

[0127] After training, an independent test set of 100 cases was used to evaluate the model performance. The model achieved a Dice coefficient of more than 96.75% on the test set, demonstrating its good segmentation performance.

[0128] S103: registering the segmented CT image with the PBV image to obtain a registered functional image;

[0129] In one embodiment, the registration is performed by combining a rigid and non-rigid algorithm to register the segmented CT image and the PBV image; a rough registration is first performed using a rigid algorithm to obtain a rough registration result, and then a non-rigid algorithm is used to perform local shape constraints on the rough registration result to obtain a registered functional image.

[0130] Optionally, the registration further includes fine adjustment, and after performing local shape constraints by a non-rigid algorithm, the local area is precisely aligned by a block matching algorithm to obtain a registered functional image;

[0131] Optionally, the rigid algorithm calculates rigid transformation parameters between the segmented CT image and the PBV image and performs rough registration based on the rigid transformation parameters to obtain a rough registration result;

[0132] Optionally, the rigid algorithm further includes parameter optimization, optimizing the rigid transformation parameters by gradient descent to obtain optimized rigid transformation parameters, and performing rough registration based on the optimized rigid transformation parameters to obtain a rough registration result;

[0133] Optionally, the non-rigid algorithm performs local deformation constraints through a B-spline deformation model.

[0134] In one embodiment, the finely adjusted area includes: lung lobe boundaries and vascular structures.

[0135] In one embodiment, the registered functional image is quantitatively evaluated through the structural similarity index and the registered mutual information value. When the quantitative evaluation result is unqualified, re-registration or manual adjustment is performed until the quantitative evaluation result is qualified.

[0136] In a specific embodiment, the segmented CT image is registered with the PBV image to obtain a registered functional image. The registration process combines rigid and non-rigid registration algorithms, specifically using a B-Spline deformation model and mutual information as a registration method for similarity measurement. The main steps of the registration algorithm include:

[0137] a) Initial alignment: First, a rough rigid registration is performed to correct the overall position and orientation differences between CT and PBV images. Mutual Information is used as the similarity metric, and the gradient descent method is used to optimize the rigid transformation parameters, including translation, rotation, and scaling. To improve efficiency, a multi-resolution strategy is adopted to gradually transition from low resolution to high resolution. This step can effectively eliminate large-scale deviations caused by scanning position differences and respiratory motion.

[0138] b) Non-rigid deformation: Based on the initial alignment, the B-spline deformation model is used for non-rigid registration to compensate for local deformation differences. The control point grid is evenly distributed on the image, and non-rigid deformation is achieved by optimizing the position of the control points. Normalized Cross Correlation is used as the local similarity measure, and the objective function is constructed in combination with the image gradient information. The L-BFGS algorithm is used to iteratively optimize the control point position and gradually refine the deformation field. To prevent excessive deformation, the elastic energy constraint term is introduced.

[0139] c) Fine-tuning and quality control: Perform local fine-tuning and registration quality assessment. Use a block matching algorithm to perform precise alignment in local areas, paying special attention to key areas such as lobe boundaries and vascular structures. Apply discrete optimization techniques, such as the graph cut algorithm, to further optimize the smoothness and consistency of the deformation field. The registration quality is quantitatively evaluated using the structural similarity index (SSIM) and the mutual information value after registration. For areas that do not meet the quality standards, they are marked and prompted to indicate that manual intervention may be required. Finally, high-quality registration results and corresponding deformation fields are output, providing a reliable foundation for subsequent functional information mapping.

[0140] In a specific embodiment, the segmented lung mask is applied to the CT image and the PBV image to extract the lung area. Then, the CT image and the PBV image are registered using a non-rigid registration algorithm. This embodiment uses a B-spline deformation model and mutual information as a registration method for similarity measurement. The main steps of the registration algorithm include:

[0141] a) Initialization:

[0142] First, an affine transformation is used to align the overall position and orientation of the two images. Affine transformation includes translation, rotation, and scaling. The affine transformation parameters are optimized by minimizing the mutual information between the two images.

[0143] b) Non-rigid deformation:

[0144] Deformation model: B-spline free deformation model is adopted;

[0145] Control point grid: a 10×10×10 control point grid is evenly distributed on the image;

[0146] Similarity measure: Local normalized cross correlation (LNCC) is used as the similarity measure;

[0147] Regularization term: Add bending energy term as regularization to prevent excessive deformation;

[0148] c) B-spline deformation model:

[0149] The deformation field model is established using a three-dimensional B-spline function. A grid of control points is evenly distributed on the image, and non-rigid deformation is achieved by optimizing the positions of the control points. The spacing of the control points varies with the resolution level, gradually decreasing from coarse to fine.

[0150] d) Optimization:

[0151] The mutual information similarity metric is optimized using gradient descent. At each resolution level, the positions of the B-spline control points are iteratively optimized until a convergence condition or a maximum number of iterations is reached.

[0152] e) Interpolation:

[0153] The deformed image is calculated using cubic spline interpolation to ensure the smoothness of image deformation.

[0154] The mutual information value and structural similarity index (SSIM) of the registered images are used to evaluate the registration quality. In this embodiment, the SSIM values ​​of the registered CT and PBV images reach above 0.85, indicating that the registration quality is good.

[0155] S104: extracting functions based on the registered functional images to obtain functional information;

[0156] In one embodiment, the functional information includes local blood flow, blood volume, and mean transit time.

[0157] In one embodiment, the functional information further includes a perfusion abnormality region, and the perfusion abnormality region is obtained by performing perfusion region detection on the registered functional image using an edge detection algorithm.

[0158] In a specific embodiment, functional information of the lung region is extracted based on the registered functional image. The main extracted functional parameters include:

[0159] Local blood flow (unit: ml / 100ml / min);

[0160] Blood volume (unit: ml / 100ml);

[0161] These parameters are obtained by analyzing and calculating the pixel values ​​of PBV images, reflecting the blood perfusion status of different lung regions.

[0162] In a specific embodiment, based on the registration results, functional information corresponding to the lung area is extracted from the PBV image. The main extracted functional parameters include: blood flow, blood volume, mean transit time, etc. These parameters are obtained by analyzing and calculating the pixel values ​​of the PBV image, reflecting the blood perfusion status of different lung areas. The specific calculation method is as follows:

[0163] a) Blood flow:

[0164] The maximum slope method is used for calculation. At each voxel, the rising segment of the PBV time-density curve is analyzed and its maximum slope is calculated. The local blood flow is proportional to the maximum slope.

[0165] b) Blood volume:

[0166] It is calculated by analyzing the ratio of the area under the PBV time-density curve to the area under the arterial input function (AIF) curve. First, the main pulmonary artery is automatically selected as the AIF, and then the area under the curve is calculated using a numerical integration method.

[0167] c) Average transit time:

[0168] According to the central volume principle, mean transit time is equal to blood volume divided by blood flow.

[0169] S105: Divide the lung function areas according to the functional information to obtain a division result.

[0170] In one embodiment, the regional divisions include a perfusion defect area, a reduced perfusion area, and a healthy area.

[0171] In one embodiment, the division is to input the functional information and the functional image into a classifier for division to obtain a division result.

[0172] In one embodiment, the process of dividing by classifier is:

[0173] Obtain functional images and functional information;

[0174] Performing feature conversion on the functional information to obtain functional features;

[0175] Extract spatial distribution features based on functional images;

[0176] The functional features and spatial distribution features are input into the classifier for classification to obtain the classification results.

[0177] In one embodiment, the abnormal perfusion area includes a perfusion defect or decreased perfusion.

[0178] In one embodiment, in the process of dividing by classification, the functional information includes one or more of the following: local blood flow, blood volume, average transit time, and abnormal perfusion area. The abnormal perfusion area is subjected to feature extraction by edge detection algorithm to obtain the imaging feature of the abnormal perfusion area (one of the functional features).

[0179] In one embodiment, the process of dividing by a classifier also includes region elimination, and a result of the removed region is obtained after region elimination, and evaluation is performed based on the result of the removed region; the region elimination is to perform morphological processing on the region in the division result to remove isolated small regions to obtain the result of the removed region.

[0180] In one embodiment, the classifier includes one or more of the following: random forest, decision tree, support vector machine, extreme learning machine, perceptron vector machine, convolutional neural network, residual network, and hole convolutional neural network.

[0181] In another embodiment, the division is obtained by comparing the local blood flow with a preset threshold value;

[0182] Optionally, the preset threshold of local blood flow in the perfusion defect area is smaller than that in the perfusion reduced area, and the preset threshold of local blood flow in the perfusion reduced area is smaller than that in the healthy area;

[0183] In one embodiment, when the local blood flow is less than 10 ml / 100 ml / min, it is determined as a perfusion defect area, and when the local blood flow is greater than or equal to 30 ml / 100 ml / min, it is determined as a healthy area;

[0184] In another embodiment, the division is first performed by comparing the local blood flow with a preset threshold to obtain a divided area, and then the divided area and functional information are input into a classifier to perform fine regional division to obtain a division result.

[0185] In a specific embodiment, the extracted functional information is analyzed and divided, and the lung area is divided into a perfusion defect (Defect Area), a perfusion reduction (Reduced Area) and a healthy area (Healthy Area). The division method adopts a method based on a combination of adaptive threshold and machine learning:

[0186] a) First, set the initial threshold based on the statistical data of the normal population;

[0187] Perfusion defect: local blood flow <10ml / 100ml / min;

[0188] Decreased perfusion: 10ml / 100ml / min≤local blood flow<30ml / 100ml / min;

[0189] Healthy area: local blood flow ≥30ml / 100ml / min;

[0190] b) Then, the preliminary segmentation results are optimized using a random forest model (RF) classifier. The feature inputs of the RF classifier include local blood flow, blood volume, and mean transit time, as well as the spatial distribution characteristics of these parameters;

[0191] c) Finally, morphological post-processing operations are applied to remove isolated small areas and obtain the final functional area division results.

[0192] The specific steps include:

[0193] Apply an opening operation (spherical structure element with a radius of 2 voxels) to the binary mask of each category;

[0194] Apply a closing operation (spherical structure element with a radius of 3 voxels) to the processed Mask;

[0195] Remove connected regions with less than 50 voxels;

[0196] The partitioning results are visualized by color coding, such as Figure 5 As shown, red indicates perfusion defect areas, blue indicates reduced perfusion areas, and green indicates healthy areas.

[0197] In a specific embodiment, in order to further improve the accuracy of diagnosis, especially for the identification of pulmonary vascular disease, this embodiment further includes the following steps:

[0198] ROI segmentation:

[0199] The whole lung image is segmented into regions of interest (ROI) to obtain a segmented whole lung image containing 36 sub-regions. The specific segmentation method is as follows:

[0200] The lungs were divided into left and right lungs. Each lung was divided into three equal regions: upper, middle, and lower in the vertical direction. Each vertical region was divided into five equal regions: front, back, inside, outside, and center in the horizontal direction. The center region was further divided into two regions: near the hilum and far from the hilum. Thus, each lung was divided into 18 sub-regions, and the whole lung had a total of 36 sub-regions. The segmented regions of interest were subjected to image registration, functional information extraction, and regional division.

[0201] In one embodiment, when dividing the lung functional areas, the lungs include one or more of the following: left lung, right lung, left and right lung.

[0202] In one embodiment, the overall process of the present invention is as follows Figure 5 As shown, dual-energy CT scan data of the lungs of the subject are obtained, including lung CT images and lung PBV images, the lung CT images are segmented to obtain lung regions (left and right lungs or left lung or right lung), the segmented CT images are registered with the PBV to obtain a registered functional image (registered PBV image), functional information is extracted from the registered PBV image, and the registered PBV image is divided into lung regions according to the functional information to obtain a division result.

[0203] All steps of the automatic classification method for pulmonary blood perfusion based on dual-energy CT in the present invention are executed by a computer for auxiliary diagnosis.

[0204] The disclosed embodiments of the present invention further provide a computer program product or system, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT.

[0205] Figure 2 A schematic diagram of an automatic classification system for pulmonary blood perfusion based on dual-energy CT provided by an embodiment of the present invention specifically includes:

[0206] Acquisition module: acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images;

[0207] Segmentation module: segmenting the CT image into lungs to obtain a segmented CT image;

[0208] Registration module: registering the segmented CT image with the PBV image to obtain a registered functional image;

[0209] Functional module: extracting functions based on the registered functional images to obtain functional information;

[0210] Division module: divides the lung functional areas according to the functional information to obtain the division results.

[0211] Figure 3 A schematic diagram of an automatic classification device for pulmonary blood perfusion based on dual-energy CT provided by an embodiment of the present invention specifically includes:

[0212] A memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, any one of the above-mentioned automatic classification methods for pulmonary blood perfusion based on dual-energy CT is performed.

[0213] The disclosed embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it is any one of the above-mentioned automatic classification methods of pulmonary blood perfusion based on dual-energy CT.

[0214] The embodiment of the present invention further provides a pulmonary blood perfusion assessment method based on dual-energy CT, comprising:

[0215] Acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images;

[0216] The segmentation results were obtained based on the above-mentioned automatic classification method of pulmonary blood perfusion based on dual-energy CT;

[0217] A pulmonary blood perfusion assessment is performed based on the segmentation result to obtain an assessment result.

[0218] In a specific embodiment, a lung perfusion assessment report is generated based on the functional area division results to assist doctors in making lung perfusion judgments. The report includes:

[0219] a) Left and right lung volumes (Volume);

[0220] b) The volume of each functional area (perfusion defect, decreased perfusion, healthy area) and its percentage of the whole lung;

[0221] c) Average blood flow and blood volume in each functional area;

[0222] d) 2D visualization image of functional area distribution;

[0223] e) comparison with normal reference values;

[0224] f) Longitudinal comparison (if there are previous examination data).

[0225] In another embodiment, a lung perfusion assessment report is generated based on the functional area division results and feature analysis results to assist doctors in making lung perfusion judgments. The report includes:

[0226] a) Left and right lung volumes;

[0227] b) The volume of each functional area (perfusion defect, decreased perfusion, healthy area) and its percentage of the whole lung;

[0228] c) Average blood flow, blood volume and average transit time in each functional area;

[0229] d) 3D visualization image of functional area distribution;

[0230] e) comparison with normal reference values;

[0231] f) longitudinal comparison (if there are previous inspection data);

[0232] g) Disease classification results and their confidence based on feature analysis;

[0233] h) Treatment recommendations: Based on the classification results, specific treatment plans are recommended for any lung disease (including typical and atypical);

[0234] The report adopts a structured format, including numerical results, charts and 2D visualization, allowing doctors to zoom and slice to view the distribution of functional areas.

[0235] An embodiment of the present invention provides a pulmonary blood perfusion assessment device based on dual-energy CT, including a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-mentioned pulmonary blood perfusion assessment method based on dual-energy CT.

[0236] An embodiment of the present invention provides a computer program product, including a computer program or instructions, wherein the computer program or instructions are executed by a processor to implement the above-mentioned dual-energy CT-based pulmonary blood perfusion assessment method.

[0237] An embodiment of the present invention provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-mentioned dual-energy CT-based pulmonary blood perfusion assessment method.

[0238] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of the method relative to the default setting. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units. A person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above-mentioned embodiments may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0239] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned medium storage can be a read-only memory, a disk or an optical disk, etc.

[0240] The above is a detailed introduction to a computer device provided by the present invention. For a person skilled in the art, according to the concept of the embodiments of the present invention, there may be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. An automatic classification method for pulmonary blood perfusion based on dual-energy CT, characterized in that: include: Acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images; Performing lung segmentation on the CT image to obtain a segmented CT image; Registering the segmented CT image with the PBV image to obtain a registered functional image; Performing function extraction based on the registered functional image to obtain functional information; The lung functional area is divided according to the functional information to obtain a division result.

2. The automatic classification method of pulmonary blood perfusion based on dual-energy CT according to claim 1, characterized in that: The regional division includes perfusion defect area, perfusion reduction area, and healthy area; Optionally, the division is to input the functional information and the functional image into a classifier for division to obtain a division result; Optionally, the functional information includes local blood flow, blood volume, and mean transit time; Optionally, the process of dividing by a classifier is: Obtain functional images and functional information; Performing feature conversion on the functional information to obtain functional features; Extract spatial distribution features based on functional images; Input the functional features and spatial distribution features into the classifier for classification to obtain the classification results; Optionally, the process of dividing by the classifier further includes region elimination, and after the region elimination, a result of the removed region is obtained, and evaluation is performed based on the result of the removed region; The region elimination is to perform morphological processing on the region in the division result to remove isolated small regions to obtain the removed region result; Optionally, the classifier includes one or more of the following: random forest, decision tree, support vector machine, extreme learning machine, perceptron vector machine, convolutional neural network, residual network, and hole convolutional neural network.

3. The automatic classification method of pulmonary blood perfusion based on dual-energy CT according to claim 1, characterized in that: The division is obtained by comparing the local blood flow with a preset threshold; optionally, the preset threshold of the local blood flow in the perfusion defect area is smaller than that in the perfusion reduction area, and the preset threshold of the local blood flow in the perfusion reduction area is smaller than that in the healthy area; Optionally, when the local blood flow is less than 10 ml / 100 ml / min, it is determined as a perfusion defect area, and when the local blood flow is greater than or equal to 30 ml / 100 ml / min, it is determined as a healthy area; Optionally, the division is first performed by comparing the local blood flow with a preset threshold to obtain a divided area, and then the divided area and functional information are input into a classifier to perform fine regional division to obtain a division result.

4. The automatic classification method of pulmonary blood perfusion based on dual-energy CT according to claim 1, characterized in that: The registration is performed by combining a rigid and non-rigid algorithm to register the segmented CT image and the PBV image; a rigid algorithm is first used to perform a rough registration to obtain a rough registration result, and then a non-rigid algorithm is used to perform local shape constraints on the rough registration result to obtain a registered functional image; Optionally, the registration further includes fine adjustment, and after performing local shape constraints by a non-rigid algorithm, the local area is precisely aligned by a block matching algorithm to obtain a registered functional image; Optionally, the rigid algorithm calculates rigid transformation parameters between the segmented CT image and the PBV image and performs rough registration based on the rigid transformation parameters to obtain a rough registration result; Optionally, the rigid algorithm further includes parameter optimization, optimizing the rigid transformation parameters by gradient descent to obtain optimized rigid transformation parameters, and performing rough registration based on the optimized rigid transformation parameters to obtain a rough registration result; Optionally, the non-rigid algorithm performs local deformation constraints through a B-spline deformation model; Optionally, the finely adjusted regions include: lung lobe boundaries, vascular structures; Optionally, the registered functional image is quantitatively evaluated through a structural similarity index and a registered mutual information value. When the quantitative evaluation result is unqualified, re-registration or manual adjustment is performed until the quantitative evaluation result is qualified.

5. The automatic classification method of pulmonary blood perfusion based on dual-energy CT according to claim 1, characterized in that: The segmentation is performed by a pre-trained segmentation model to obtain a segmentation result, and the segmentation result includes a left lung and a right lung; Optionally, the segmentation model includes one or more of the following: U-Net, FCN, SegNet, DeepLabV3+, ENet; Optionally, the segmentation model includes an encoder and a decoder, the encoder consists of N residual blocks, N is a natural number greater than 1, the input data is extracted through the N residual blocks connected in series in the encoder to obtain multi-scale features, the features of the Nth residual block are input to the decoder, and the first N-1 multi-scale features are input to the decoder through jump connections to fuse features of different scales to obtain decoding features, and then the decoding features are weighted by the attention mechanism to obtain important areas, and the important areas are segmented to obtain segmentation results.

6. The automatic classification method of pulmonary blood perfusion based on dual-energy CT according to claim 5, characterized in that: The method further includes data preprocessing, performing data preprocessing on the CT image to obtain a processed CT image, and segmenting the processed image to obtain a segmented CT image; Optionally, the data preprocessing includes pixel normalization, size unification, image noise reduction, content highlighting, and data enhancement; Optionally, the content highlights lung tissue by window width and window level; Optionally, the image denoising removes image noise by Gaussian filtering; Optionally, the segmentation further includes data post-processing, performing data post-processing on the segmented CT image to obtain a post-processed CT image, and registering the post-processed CT image with the PBV image; The post-processing is to perform morphological processing on the segmented CT image to obtain a post-processed CT image; Optionally, the morphological processing includes opening and closing operations, removing isolated areas, and filling holes; Optionally, the post-processing further includes edge optimization, performing edge optimization on the morphologically processed CT image to obtain a post-processed CT image; Optionally, the edge optimization is performed by conditional random fields.

7. The method for evaluating pulmonary blood perfusion based on dual-energy CT according to claim 1, characterized in that: Acquiring dual-energy CT scan data of the lungs of the subject to be tested, wherein the dual-energy CT scan data includes CT images and PBV images; Obtaining a classification result based on the automatic classification method for pulmonary blood perfusion based on dual-energy CT according to any one of claims 1 to 6; A pulmonary blood perfusion assessment is performed based on the segmentation result to obtain an assessment result.

8. A computer product comprising a computer program or instructions, characterized in that The computer program or instructions are executed by the processor to implement the automatic classification method of pulmonary blood perfusion based on dual-energy CT as described in any one of claims 1-6 or the pulmonary blood perfusion assessment method based on dual-energy CT as described in claim 7.

9. A computer device comprising a memory, a processor and a computer program or instruction stored in the memory, characterized in that: The computer program or instructions are executed by the processor to implement the automatic classification method of pulmonary blood perfusion based on dual-energy CT as described in any one of claims 1-6 or the pulmonary blood perfusion assessment method based on dual-energy CT as described in claim 7.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: The computer program or instructions are executed by the processor to implement the automatic classification method of pulmonary blood perfusion based on dual-energy CT as described in any one of claims 1-6 or the pulmonary blood perfusion assessment method based on dual-energy CT as described in claim 7.

Images