System for assessing pulmonary ventilation function based on ct images

Abstract

The application relates to the technical field of medical image processing, and particularly discloses a lung ventilation function evaluation system based on CT images, which adopts deep learning-based image processing technology to analyze lung CT images at the end of inspiration and expiration of a patient, so as to extract semantic feature representations of the lung CT images, and then, through fine-grained semantic interaction and semantic compensation of the lung CT image features at the end of inspiration and expiration, accurate characterization of a lung deformation field is obtained, and the lung capacity of the patient is intelligently estimated. In this way, non-invasive and accurate evaluation of lung function can be realized, and the method is particularly suitable for a patient group with poor physical conditions or unable to complete standard breathing tests as required.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and more specifically, to a pulmonary ventilation function assessment system based on CT images. Background Technology

[0002] Pulmonary ventilation function is a crucial indicator for assessing respiratory health and disease progression. Vital capacity (VC) is a key parameter measuring lung function, referring to the maximum volume of air that can be exhaled after a maximal inhalation. Accurate assessment of VC is particularly important for the diagnosis and monitoring of respiratory diseases such as chronic obstructive pulmonary disease (COPD) and asthma. Traditionally, VC relies on devices like spirometers to measure changes in respiratory flow and volume under specific conditions to assess parameters such as VC, forced expiratory volume in one second (FEV1), and forced vital capacity (FVC). However, this method often requires the patient to perform specific breathing movements, which may be difficult for some frail patients or those with severe respiratory illnesses.

[0003] In recent years, lung CT (Computed Tomography) images have been widely used as a non-invasive diagnostic tool to assess abnormalities in lung structure and function. High-resolution images obtained through CT scans can provide detailed information about lung tissue structure, playing an irreplaceable role in disease detection, staging, and evaluation of treatment effectiveness. However, traditional CT image analysis has primarily focused on visualizing anatomical structures and identifying lesion areas, failing to fully utilize the rich information contained in CT images to quantify functional changes in the lungs.

[0004] Therefore, a CT image-based system for assessing pulmonary ventilation function is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a CT image-based pulmonary ventilation function assessment system. This system employs deep learning-based image processing technology to analyze lung CT images taken at the end of inspiration and end of expiration to extract semantic feature representations of the lung CT images. Furthermore, by performing fine-grained semantic interaction and semantic compensation on the lung CT image features at the end of inspiration and end of expiration, an accurate representation of the lung deformation field is obtained, thereby intelligently estimating the patient's vital capacity. This approach enables non-invasive and accurate assessment of lung function, and is particularly suitable for patients with poor physical condition or those unable to complete standard respiratory tests as required.

[0006] Accordingly, according to one aspect of this application, a CT image-based pulmonary ventilation function assessment system is provided, comprising:

[0007] Lung CT image receiving module, used to receive a collection of lung CT images for an entire respiratory cycle;

[0008] The end-expiratory lung CT image extraction module is used to extract end-inspiratory lung CT images and end-expiratory lung CT images from the set of lung CT images;

[0009] The lung CT image feature extraction module is used to extract image features from the lung CT image at the end of inspiration and the lung CT image at the end of expiration to obtain the semantic coding feature map of the lung CT image at the end of inspiration and the semantic coding feature map of the lung CT image at the end of expiration.

[0010] The global interactive coding module is used to perform fine-grained global interactive coding based on text semantic compensation on the semantic coding feature map of the lung CT image at the end of inspiration and the semantic coding feature map of the lung CT image at the end of expiration to obtain the semantic coding feature map of the lung CT deformation field at the end of inspiration and the end of expiration.

[0011] The vital capacity estimation module is used to generate vital capacity estimates based on the semantic coding feature map of lung CT deformation field at the end of inspiration and end of expiration.

[0012] Compared with existing technologies, the CT image-based pulmonary ventilation function assessment system provided in this application employs deep learning-based image processing technology to analyze lung CT images taken at the end of inspiration and end of expiration to extract semantic feature representations of the lung CT images. Furthermore, by performing fine-grained semantic interaction and semantic compensation on the lung CT image features at the end of inspiration and end of expiration, an accurate representation of the lung deformation field is obtained, which is then used to intelligently estimate the patient's vital capacity. This approach enables non-invasive and accurate assessment of lung function, and is particularly suitable for patients with poor physical condition or those unable to complete standard respiratory tests as required. Attached Figure Description

[0013] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 This is a block diagram of a CT image-based lung ventilation function assessment system according to an embodiment of this application.

[0015] Figure 2 This is a schematic diagram of data flow in a CT image-based lung ventilation function assessment system according to an embodiment of this application.

[0016] Figure 3 This is a block diagram of the global interactive coding module in a CT image-based pulmonary ventilation function assessment system according to an embodiment of this application.

[0017] Figure 4 This is a block diagram of a semantic compensation enhancement aggregation unit in a CT image-based lung ventilation function assessment system according to an embodiment of this application. Detailed Implementation

[0018] The embodiments of this application will now be described in more detail with reference to the accompanying drawings, and the above and other objects, features, and advantages of this application will become more apparent. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0019] In view of the technical problems described in the background section, this application proposes a pulmonary ventilation function assessment system based on CT images. Figure 1 This is a block diagram of a CT image-based lung ventilation function assessment system according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a CT image-based pulmonary ventilation function assessment system according to an embodiment of this application. Figure 1 and Figure 2 As shown, the CT image-based lung ventilation function assessment system 100 according to an embodiment of this application includes: a lung CT image receiving module 110, used to receive a set of lung CT images for an entire respiratory cycle; a lung CT image extraction module 120, used to extract end-inspiratory lung CT images and end-expiratory lung CT images from the set of lung CT images; and a lung CT image feature extraction module 130, used to extract image features from the end-inspiratory lung CT images and the end-expiratory lung CT images respectively to obtain... The system obtains semantic coding feature maps of lung CT images at the end of inspiration and at the end of expiration; a global interactive coding module 140 is used to perform fine-grained global interactive coding based on text semantic compensation on the semantic coding feature maps of lung CT images at the end of inspiration and at the end of expiration to obtain semantic coding feature maps of lung CT deformation field at the end of inspiration and at the end of expiration; a vital capacity estimation module 150 is used to generate vital capacity estimates based on the semantic coding feature maps of lung CT deformation field at the end of inspiration and at the end of expiration.

[0020] In the aforementioned CT image-based pulmonary ventilation function assessment system 100, the lung CT image receiving module 110 is used to receive a collection of lung CT images for an entire respiratory cycle. It should be understood that the lungs are a dynamic organ, undergoing significant volume and morphological changes during respiration. A complete respiratory cycle includes the process from the start of inspiration to maximal inspiration (end of inspiration), then to the start of expiration and maximal expiration (end of expiration). By collecting lung CT image data throughout the entire respiratory cycle, the changes in lung volume and internal structure at different respiratory stages can be fully captured, thereby providing comprehensive information for pulmonary function assessment.

[0021] Specifically, when a patient needs a lung CT scan, to obtain high-quality image data, it is essential to first ensure that the patient maintains a stable breathing pattern during the scan. This allows for accurate extraction of keyframes at the end of inspiration and end of expiration. Therefore, before the actual scan begins, medical staff need to provide detailed instructions to the patient, practicing breathing techniques such as deep or shallow breathing, to help the patient familiarize themselves with the upcoming lung CT scan procedure. This preparation helps reduce image artifacts caused by irregular breathing, thereby improving the reliability of the final diagnostic results.

[0022] Next, the operating parameters of the CT scanner need to be set according to specific clinical needs. For cases where it's necessary to capture changes in lung morphology throughout a complete respiratory cycle, respiratory gating is crucial. Respiratory gating allows the CT scanner to automatically adjust the scan timing based on the patient's respiratory status, ensuring that each scan occurs during the same or similar respiratory phase. Additionally, a multi-stage acquisition strategy can be selected, which involves rapidly and continuously capturing multiple tomographic images within a single respiratory cycle. This method covers a wider range of anatomical structures while retaining sufficient detail for subsequent analysis. Spiral CT or volumetric CT scans are preferred due to their ability to acquire a large number of high-resolution images in a short time.

[0023] During operation, a CT scanner assigns a timestamp to each generated tomographic image, indicating the specific moment the image was acquired. After the scan is complete, all acquired CT images are organized into an ordered sequence reflecting changes in lung morphology throughout the respiratory cycle. However, simply arranging these images is insufficient for further analysis. Instead, a series of preprocessing operations must be performed, including but not limited to geometric correction, contrast adjustment, and noise suppression, to ensure that each image clearly displays the lung structure and maintains good consistency among them.

[0024] While automation tools have greatly simplified the process, in some cases, a final quality check by professionals is still necessary to confirm that the image truly meets the expected standards. If any issues are found, such as blurry images or excessive motion artifacts, it may be necessary to reconsider and select other more suitable candidate images. Only images that have undergone rigorous review will be deemed suitable for the next step of functional evaluation.

[0025] Once the final image set is determined, the next step is to integrate the image data into a CT image-based pulmonary ventilation function assessment system. This integration process is not simply a file transfer, but involves a series of complex data conversions and format adaptations. For example, the original DICOM format images may need to be converted to a format more suitable for deep learning algorithms; in addition, it is necessary to ensure that the relative positional relationships between the images are correctly maintained so that subsequent spatial transformation operations can proceed smoothly.

[0026] In summary, acquiring a collection of lung CT images spanning an entire respiratory cycle is a highly specialized and technology-intensive process. It requires close collaboration among interdisciplinary teams, integrating modern medical imaging techniques and computer science knowledge to successfully construct a stable and reliable image acquisition and processing system. The high-quality image data obtained in this way lays a solid foundation for CT-based pulmonary ventilation function assessment systems, enabling a deeper understanding of the dynamic changes in the lungs and providing strong support for clinical diagnosis and treatment decisions.

[0027] In the aforementioned CT image-based pulmonary ventilation function assessment system 100, the end-expiratory lung CT image extraction module 120 is used to extract end-inspiratory and end-expiratory lung CT images from the set of lung CT images. It should be understood that to accurately estimate a patient's vital capacity parameters, it is necessary to understand the morphological changes of the lungs under extreme respiratory conditions. Considering that end-inspiratory and end-expiratory phases represent two extreme states within the respiratory cycle, corresponding to the maximum and minimum lung volumes respectively, the difference in lung volume between the two directly reflects the patient's vital capacity. Therefore, this application further extracts end-inspiratory and end-expiratory lung CT images from the set of lung CT images, and by comparing and analyzing the CT images under these two extreme states, estimates the range of lung deformation and its impact on gas exchange, thereby estimating the patient's vital capacity.

[0028] In practice, in order to accurately extract lung CT images at the end of inspiration and at the end of expiration, it is also necessary to simultaneously acquire the patient's respiratory signal during the CT scan and match it with the timestamp of the CT image to identify the lung CT images representing the end of inspiration and at the end of expiration.

[0029] Specifically, as the CT scan begins, the patient breathes as previously instructed, while the medical team uses specially designed equipment to record the patient's breathing patterns. This equipment can be a simple breathing belt that senses respiratory movements by monitoring the expansion and contraction of the chest or abdomen; or a more complex flow sensor that directly measures the airflow into and out of the respiratory tract. Regardless of the method used, the goal is to obtain a clear, continuous respiratory waveform that reflects the changing characteristics of each phase of the respiratory cycle.

[0030] Next, the acquired respiratory signals are preprocessed. This step includes, but is not limited to, filtering to remove noise, correcting baseline drift, and detecting and labeling key events of the respiratory cycle (such as the onset and end of inspiration, the onset and end of expiration). These key events are crucial for subsequently selecting the correct CT images. The preprocessed respiratory signals provide a timeframe to guide the selection of the most suitable images from the CT image set.

[0031] During a CT scan, each generated tomographic image is assigned a timestamp, representing the exact moment the image was acquired. By matching these CT image timestamps with the pre-processed respiratory signal timeline, the CT images corresponding to the end of inspiration and end of expiration can be accurately identified. Ideally, two specific time points need to be found: the instant at the end of inspiration and the instant at the end of expiration. However, in practice, due to potentially irregular breathing rhythms and the influence of CT scan speed, it is difficult to capture these two instants with absolute precision. Therefore, in practice, the CT images closest to these two time points are often sought, and sometimes multiple images covering a short time period are selected and then averaged to obtain more stable and representative results. This method can effectively reduce errors caused by irregular breathing and variations in scan speed, thereby improving image quality and diagnostic accuracy.

[0032] Once candidate CT images are identified, the next step is to evaluate their quality. This step directly impacts the accuracy of the final analysis. Specifically, first, the images need to be checked for obvious artifacts, such as blurring or distortion caused by respiratory movements. Furthermore, it must be confirmed that the selected images accurately reflect the expected changes in lung morphology—for example, the lung tissue should appear more full at the end of inspiration and relatively contracted at the end of expiration. For this purpose, automated algorithms may be used to perform consistency checks, comparing the similarities and differences between images from different time points to verify that they conform to the expected morphological change patterns. This evaluation process ensures that the CT images used are of high quality, providing a reliable basis for subsequent analysis. This not only helps improve the accuracy of the analysis but also ensures that the patient's lung condition can be accurately assessed, providing strong support for clinical decision-making.

[0033] In addition to automation and technological advancements, the professional knowledge of clinicians is equally indispensable. With their rich experience and intuition, they can manually adjust the output of the automated system when necessary, ensuring that the final selected images are both scientifically sound and consistent with clinical realities. This human-machine collaboration not only improves work efficiency but also guarantees diagnostic quality.

[0034] In summary, extracting end-inspiratory and end-expiratory images from a collection of lung CT images is a multidisciplinary task, combining techniques and methods from multiple fields such as biosignal processing, medical imaging, and computer science. Through meticulous analysis of respiratory signals, precise matching with CT image timestamps, and rigorous control of image quality, high-quality representative images can be obtained, providing a solid foundation for subsequent pulmonary function assessment.

[0035] In the aforementioned CT image-based lung ventilation function assessment system 100, the lung CT image feature extraction module 130 is used to extract image features from the end-inspiratory lung CT image and the end-expiratory lung CT image respectively to obtain a semantic coding feature map of the end-inspiratory lung CT image and a semantic coding feature map of the end-expiratory lung CT image. In a specific example of this application, the lung CT image feature extraction module 130 is used to: input the end-inspiratory lung CT image and the end-expiratory lung CT image respectively into a MobileNet-based lung CT image feature extractor to obtain a semantic coding feature map of the end-inspiratory lung CT image and a semantic coding feature map of the end-expiratory lung CT image. That is, in order to achieve efficient feature extraction from lung CT images, this application uses MobileNet to process the end-inspiratory lung CT image and the end-expiratory lung CT image respectively. Those skilled in the art will know that MobileNet, as a lightweight deep learning model, reduces computational resource consumption while maintaining high accuracy by introducing depthwise separable convolution operations, making it ideal for processing high-resolution CT image data. In the technical solution of this application, the use of the MobileNet model can effectively extract key lung structural features, such as alveolar expansion state, lung lobe shape, and size, from the end-inspiratory and end-expiratory lung CT images, thereby providing accurate image feature input for subsequent vital capacity estimation.

[0036] In a specific example of this application, the MobileNet includes an input layer, an initial layer, multiple depthwise separable convolutional blocks, an average pooling layer, and an output layer. The initial layer includes a standard Conv2D convolutional layer with a 3x3 filter size and a stride of 2, followed by a batch normalization layer and a ReLU activation function. Each depthwise separable convolutional block includes a depthwise convolutional layer and a pointwise convolutional layer. The depthwise convolutional layer applies a 3x3 convolutional kernel to each input channel individually, maintaining the same number of channels, and is followed by a batch normalization layer and a ReLU activation function. The pointwise convolutional layer uses a 1x1 convolutional kernel to convert the number of channels from input channels to output channels, and is also followed by a batch normalization layer and a ReLU activation function.

[0037] In the aforementioned CT image-based pulmonary ventilation function assessment system 100, the global interactive coding module 140 is used to perform fine-grained global interactive coding based on text semantic compensation on the semantic coding feature maps of the end-inspiratory lung CT images and the end-expiratory lung CT images to obtain a semantic coding feature map of the lung CT deformation field from end-inspiratory to end-expiratory. It should be understood that the changes in lung morphology between the end-inspiratory and end-expiratory phases directly reflect the patient's vital capacity. Therefore, this application further captures and quantifies the morphological changes of the lungs during the respiratory cycle by interactively analyzing the semantic coding feature maps of the end-inspiratory and end-expiratory lung CT images, achieving a deeper understanding of the lung deformation field, thereby providing important information for assessing the patient's vital capacity. Based on this, in order to improve the analysis accuracy of the lung deformation field, this application proposes a fine-grained interaction method based on text semantic compensation. This method utilizes the prior information of a large language model to perform semantic compensation of fine-grained interaction features between the semantic coding feature maps of the lung CT images at the end of inspiration and at the end of expiration, so as to ensure the generation of richer and more accurate lung deformation field feature representations, thereby improving the accuracy of vital capacity estimation.

[0038] Figure 3 This is a block diagram of the global interactive coding module in a CT image-based pulmonary ventilation function assessment system according to an embodiment of this application. Figure 3 As shown, the global interactive encoding module 140 includes: a local fine-grained interaction unit 141, used to extract the local shared semantic features of the semantic encoding feature map of the lung CT image at the end of inspiration and the semantic encoding feature map of the lung CT image at the end of expiration to obtain a set of fine-grained local shared semantic feature matrices of lung CT at the end of inspiration and expiration; a semantic decoding unit 142, used to perform semantic decoding on the set of fine-grained local shared semantic feature matrices of lung CT at the end of inspiration and expiration to generate a set of semantic compensation text descriptions of lung CT deformation field at the end of inspiration and expiration; and a semantic compensation enhancement aggregation unit 143, used to perform semantic compensation enhancement aggregation on the set of fine-grained local shared semantic feature matrices of lung CT at the end of inspiration and expiration based on the set of semantic compensation text descriptions of lung CT deformation field at the end of inspiration and expiration to obtain the semantic encoding feature map of lung CT deformation field at the end of inspiration and expiration.

[0039] Specifically, the local fine-grained interaction unit 141 is used to: perform feature decomposition on the semantic coding feature maps of the end-inspiratory and end-expiratory lung CT images along the channel dimensions to obtain a set of local semantic feature matrices for the end-inspiratory and end-expiratory lung CT images; input the local semantic feature matrices of the end-inspiratory and end-expiratory lung CT images corresponding to each channel dimension in the set of local semantic feature matrices for the end-inspiratory and end-expiratory lung CT images into a shared semantic information extraction module based on a Siamese network structure to obtain a set of fine-grained local shared semantic feature matrices for the end-inspiratory and end-expiratory lung CT images. The shared semantic information extraction module based on a Siamese network structure includes three parallel feature interaction layers, a feature cascade layer, a point convolutional layer, and an activation layer based on the sigmoid function.

[0040] Accordingly, the calculation process of the aforementioned local fine-grained interaction unit 141 can be expressed by the following formula:

[0041] Decouple(F1) = {F 11 ,F 12 ,...,F 1i ,...,F 1n}

[0042] Decouple(F2) = {F 21 ,F 22 ,...,F 2i ,...,F 2n}

[0043]

[0044] Wherein, F1 represents the semantic coding feature map of the lung CT image at the end of inspiration, F2 represents the semantic coding feature map of the lung CT image at the end of expiration, and Decouple(·) represents feature dissociation. 11 F 12 F 1i and F 1n F represents the first, second, i-th, and n-th local semantic feature matrices of the end-inspiratory lung CT image semantic coding feature map along the channel dimension, respectively, where n is the number of channels in the end-inspiratory lung CT image semantic coding feature map. 21 F 22 F 2i and F 2nThese represent the first, second, i-th, and n-th local semantic feature matrices of the end-tidal lung CT image semantic coding feature map along the channel dimension. The circle (⊙) represents addition by position, and the circle (⊙) represents multiplication by position. This indicates subtraction by position point; `concat(·;·;·)` indicates feature concatenation; `Conv` 1×1 This represents a pointwise convolution operation, sigmoid represents the sigmoid activation function, and S... i This represents the fine-grained local shared semantic feature matrix of the i-th lung CT at the end of inspiration and end of expiration.

[0045] It should be understood that in the semantic coding feature maps of the end-inspiratory and end-expiratory lung CT images, different feature channels typically carry different information content. By decomposing the feature maps into smaller local feature matrices, more detailed changes in lung structure can be captured, thus providing a foundation for subsequent fine-grained feature interaction analysis. Next, a shared semantic information extraction module based on a Siamese network structure is used to extract shared semantic information from the local semantic feature matrices of the end-inspiratory and end-expiratory lung CT images for each corresponding channel. This shared semantic information extraction module, through multi-level feature interaction processing, can effectively uncover common features between each group of local features, learning a deeper semantic correlation between lung structures in the end-inspiratory and end-expiratory lung CT images, thereby achieving fine-grained interaction between lung features in the end-inspiratory and end-expiratory phases.

[0046] Specifically, the semantic decoding unit 142 is used to: input each of the end-inspiratory-end-expiratory lung CT fine-grained local shared semantic feature matrices from the set of end-inspiratory-end-expiratory lung CT fine-grained local shared semantic feature matrices into the semantic compensation decoding module based on a large language model to obtain the set of semantic compensation text descriptions of the end-inspiratory-end-expiratory lung CT deformation field, expressed by the formula:

[0047] T i =LLM{S i}

[0048] Where LLM{·} represents a large language model, T i This represents the semantic compensation text description of the lung CT deformation field at the end of the i-th inspiratory-expiratory phase.

[0049] Here, considering that focusing only on shared semantic features during feature interaction would result in the loss of other features, such as complementary features, we further compensate for feature interaction by inputting the fine-grained local shared semantic feature matrices of each end-inspiratory-end-expiratory lung CT image into a large language model for semantic compensation decoding. This leverages the rich semantic knowledge base of the large language model to supplement and enhance the input fine-grained local shared semantic feature matrices of the end-inspiratory-end-expiratory lung CT images, generating a semantically compensated text description of the lung CT deformation field. This allows us to obtain complementary features between the end-inspiratory and end-expiratory lung CT images, in addition to shared semantic information.

[0050] Figure 4 This is a block diagram of a semantic compensation enhancement aggregation unit in a CT image-based pulmonary ventilation function assessment system according to an embodiment of this application. Figure 4 As shown, the semantic compensation enhancement aggregation unit 143 includes: a semantic encoding subunit 1431, used to input each of the end-inspiratory-end-expiratory lung CT deformation field semantic compensation text descriptions in the set of end-inspiratory-end-expiratory lung CT deformation field semantic compensation text descriptions into a semantic encoder based on a text convolutional neural network model to obtain a set of end-inspiratory-end-expiratory lung CT deformation field semantic compensation text semantic encoding feature matrices; and a semantic interaction compensation subunit 1432, used to combine the set of end-inspiratory-end-expiratory lung CT fine-grained local shared semantic feature matrices with the end-inspiratory-end-expiratory lung CT deformation field semantic compensation text semantic encoding feature matrices. The set of semantically compensated text semantic coding feature matrices for CT deformation fields, including the fine-grained local shared semantic feature matrix of lung CT at the end of inspiration and the semantically compensated text semantic coding feature matrix of lung CT deformation fields at the end of inspiration and the end of expiration, are input into the fine-grained semantic interactive compensation module to obtain the set of local semantic enhancement feature matrices of lung CT deformation fields at the end of inspiration and the end of expiration. The feature aggregation subunit 1433 is used to aggregate the set of local semantic enhancement feature matrices of lung CT deformation fields at the end of inspiration and the end of expiration along the channel dimension to obtain the semantic coding feature map of lung CT deformation fields at the end of inspiration and the end of expiration.

[0051] More specifically, the semantic interaction compensation subunit 1432 is used to: use the semantic encoding feature matrix of the lung CT deformation field semantic compensation text at the end of inspiration and end of expiration as the key matrix, and use the fine-grained local shared semantic feature matrix of the lung CT at the end of inspiration and end of expiration as the query matrix and value matrix, input the query matrix, the value matrix and the key matrix into the converter unit in the fine-grained semantic interaction compensation module to obtain the local semantic query interaction feature matrix of the lung CT deformation field at the end of inspiration and end of expiration; calculate the position-weighted sum of the local semantic query interaction feature matrix of the lung CT deformation field at the end of inspiration and end of expiration and the fine-grained local shared semantic feature matrix of the lung CT at the end of inspiration and end of expiration to obtain the local semantic enhancement feature matrix of the lung CT deformation field at the end of inspiration and end of expiration.

[0052] Accordingly, the calculation process of the semantic compensation enhancement aggregation unit 143 described above can be expressed by the following formula:

[0053] M i =TextCNN{T i}

[0054]

[0055] F f =couple{S b1 ,S b2 ,...,S bi ,...,S bn}

[0056] Where TextCNN represents a text convolutional neural network model, M i Let represent the semantic encoding feature matrix of the semantic compensation text for the lung CT deformation field at the end of inspiration and end of expiration, (·). T Represents the transpose of a matrix. This represents a matrix multiplication operation, where α and β represent different weighting parameters, and S represents the feature scale value of the semantic encoding feature matrix of the lung CT deformation field semantic compensation text at the end of inspiration and end of expiration. b1 S b2 S bi and S bn Let F represent the local semantic enhancement feature matrices of the lung CT deformation field at the end of inspiration and end of expiration, respectively. Let couple{·,·,·,·} denote feature concatenation. f This represents the semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration.

[0057] In other words, to apply the semantic compensation information contained in the generated text descriptions to the fine-grained local shared semantic feature matrices of each end-inspiratory-end-expiratory lung CT image, this application further utilizes a semantic encoder based on a text convolutional neural network model to perform semantic embedding encoding on each text description for feature alignment, achieving a unified feature dimension, thereby facilitating the effective fusion of subsequent semantic compensation information. Then, based on the converter architecture, cross-domain query attention interaction is performed between each corresponding end-inspiratory-end-expiratory lung CT deformation field semantic encoding feature matrix and the end-inspiratory-end-expiratory lung CT deformation field semantic compensation text semantic encoding feature matrix to further refine and enhance the lung deformation field features. It should be understood that the converter architecture, through the attention mechanism, can effectively identify and focus on key regions in lung CT image features and key semantic information in text descriptions, capturing subtle changes in the lungs under different respiratory states, thereby effectively enhancing the expressive power of lung deformation features and providing solid data support for the accurate assessment of vital capacity. Finally, the semantic query interaction features between cross-modal data obtained after the converter are weighted and fused with the fine-grained local shared semantic features of lung CT at the end of inspiration and end of expiration to further emphasize the local fine-grained details of lung morphological changes. The weighted and fused local feature representations are then aggregated along the channel dimension to restore the original high-dimensional feature structure, resulting in the final semantic coding feature map of lung deformation field at the end of inspiration and end of expiration.

[0058] In the aforementioned CT image-based pulmonary ventilation function assessment system 100, the vital capacity estimation module 150 is used to generate a vital capacity estimate based on the semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration. In a specific example of this application, the vital capacity estimation module 150 is used to: input the semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration into a decoder-based pulmonary ventilation function assessment module to obtain the vital capacity estimate. Specifically, the decoder adopts a convolutional neural network architecture, and performs deep learning processing on the input semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration through a series of convolutional layers, pooling layers, and activation functions to progressively extract high-level features of lung structural deformation, and maps the learned high-level features to the numerical estimate of vital capacity through a fully connected layer, thereby achieving an accurate estimate of the patient's vital capacity. In this way, it can provide strong auxiliary support for the clinical diagnosis and treatment of patients' lung function.

[0059] In summary, the CT image-based pulmonary ventilation function assessment system according to the embodiments of this application is explained. It employs deep learning-based image processing technology to analyze lung CT images taken at the end of inspiration and end of expiration to extract semantic feature representations of the lung CT images. Furthermore, by performing fine-grained semantic interaction and semantic compensation on the lung CT image features at the end of inspiration and end of expiration, an accurate representation of the lung deformation field is obtained, thereby intelligently estimating the patient's vital capacity. In this way, non-invasive and accurate assessment of lung function can be achieved, making it particularly suitable for patients with poor physical condition or those unable to complete standard respiratory tests as required.

[0060] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details of the above embodiments are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the specific details described above.

Claims

1. A CT image-based pulmonary ventilation function assessment system, characterized in that, include: Lung CT image receiving module, used to receive a collection of lung CT images for an entire respiratory cycle; The end-expiratory lung CT image extraction module is used to extract end-inspiratory lung CT images and end-expiratory lung CT images from the set of lung CT images; The lung CT image feature extraction module is used to extract image features from the lung CT image at the end of inspiration and the lung CT image at the end of expiration to obtain the semantic coding feature map of the lung CT image at the end of inspiration and the semantic coding feature map of the lung CT image at the end of expiration. The global interactive coding module is used to perform fine-grained global interactive coding based on text semantic compensation on the semantic coding feature map of the lung CT image at the end of inspiration and the semantic coding feature map of the lung CT image at the end of expiration to obtain the semantic coding feature map of the lung CT deformation field at the end of inspiration and the end of expiration. The vital capacity estimation module is used to generate vital capacity estimates based on the semantic coding feature map of lung CT deformation field at the end of inspiration and end of expiration.

2. The CT image-based pulmonary ventilation function assessment system according to claim 1, characterized in that, The lung CT image feature extraction module is used for: The end-inspiratory lung CT image and the end-expiratory lung CT image are respectively input into a MobileNet-based lung CT image feature extractor to obtain the semantic coding feature map of the end-inspiratory lung CT image and the semantic coding feature map of the end-expiratory lung CT image.

3. The CT image-based pulmonary ventilation function assessment system according to claim 2, characterized in that, The global interactive encoding module includes: Local fine-grained interaction units are used to extract the locally shared semantic features of the semantic coding feature map of the lung CT image at the end of inspiration and the semantic coding feature map of the lung CT image at the end of expiration to obtain a set of fine-grained local shared semantic feature matrices of lung CT at the end of inspiration and the end of expiration. The semantic decoding unit is used to perform semantic decoding on the set of fine-grained local shared semantic feature matrices of the lung CT at the end of inspiration and end of expiration to generate a set of semantic compensation text descriptions of the lung CT deformation field at the end of inspiration and end of expiration. The semantic compensation enhancement aggregation unit is used to perform semantic compensation enhancement aggregation on the set of fine-grained locally shared semantic feature matrices of the lung CT deformation field at the end of inspiration and end of expiration, based on the set of semantic compensation text descriptions of the lung CT deformation field at the end of inspiration and end of expiration, so as to obtain the semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration.

4. The CT image-based pulmonary ventilation function assessment system according to claim 3, characterized in that, The local fine-grained interaction unit is used for: The semantic coding feature maps of the lung CT images at the end of inhalation and at the end of exhalation are decomposed along the channel dimensions to obtain a set of local semantic feature matrices of the lung CT images at the end of inhalation and at the end of exhalation. The set of local semantic feature matrices of the lung CT images at the end of inspiration and the set of local semantic feature matrices of the lung CT images at the end of expiration are input into the shared semantic information extraction module based on the Siamese network structure to obtain the set of fine-grained local shared semantic feature matrices of the lung CT images at the end of inspiration and the end of expiration.

5. The CT image-based pulmonary ventilation function assessment system according to claim 4, characterized in that, The shared semantic information extraction module based on the Siamese network structure includes three parallel feature interaction layers, feature concatenation layers, point convolutional layers, and activation layers based on the sigmoid function.

6. The CT image-based pulmonary ventilation function assessment system according to claim 5, characterized in that, The semantic decoding unit is used for: Each of the end-inspiratory-end-expiratory lung CT fine-grained local shared semantic feature matrices in the set of end-inspiratory-end-expiratory lung CT fine-grained local shared semantic feature matrices is input into the semantic compensation decoding module based on a large language model to obtain the set of semantic compensation text descriptions of the end-inspiratory-end-expiratory lung CT deformation field.

7. The CT image-based pulmonary ventilation function assessment system according to claim 6, characterized in that, The semantic compensation and enhancement aggregation unit includes: The semantic encoding subunit is used to input each of the semantic compensation text descriptions of the lung CT deformation field at the end of inspiration and end of expiration from the set of semantic compensation text descriptions of the lung CT deformation field at the end of inspiration and end of expiration into the semantic encoder based on the text convolutional neural network model to obtain a set of semantic encoding feature matrices of the semantic compensation text of the lung CT deformation field at the end of inspiration and end of expiration. The semantic interaction compensation subunit is used to input the set of fine-grained local shared semantic feature matrices of lung CT at the end of inspiration and at the end of expiration and the set of semantic compensation text semantic coding feature matrices of lung CT deformation field at the end of inspiration and at the end of expiration into the fine-grained semantic interaction compensation module to obtain the set of local semantic enhancement feature matrices of lung CT deformation field at the end of inspiration and at the end of expiration. The feature aggregation subunit is used to aggregate the set of local semantic enhancement feature matrices of the lung CT deformation field at the end of inspiration and end of expiration along the channel dimension to obtain the semantic coding feature map of the lung CT deformation field at the end of inspiration and end of expiration.

8. The CT image-based pulmonary ventilation function assessment system according to claim 7, characterized in that, The semantic interaction compensation subunit is used for: Using the semantic encoding feature matrix of the lung CT deformation field at the end of inspiration and end of expiration as the key matrix, and the fine-grained local shared semantic feature matrix of the lung CT at the end of inspiration and end of expiration as the query matrix and value matrix, the query matrix, the value matrix and the key matrix are input into the converter unit in the fine-grained semantic interaction compensation module to obtain the local semantic query interaction feature matrix of the lung CT deformation field at the end of inspiration and end of expiration. The position-weighted sum of the local semantic query interaction feature matrix of the lung CT deformation field at the end of inspiration and the fine-grained local shared semantic feature matrix of the lung CT at the end of inspiration and the end of expiration is calculated to obtain the local semantic enhancement feature matrix of the lung CT deformation field at the end of inspiration and the end of expiration.

9. The CT image-based pulmonary ventilation function assessment system according to claim 8, characterized in that, The vital capacity estimation module is used for: The semantic encoding feature map of lung CT deformation field at the end of inspiration and end of expiration is input into the decoder-based lung ventilation function assessment module to obtain the estimated vital capacity value.

Images