Automatic evaluation system and method for laryngotracheal stenosis

Abstract

The invention relates to the technical field of automatic image analysis, and provides an automatic evaluation system and method for laryngotracheal stenosis, and the system comprises a laryngotracheal segmentation module which is configured to carry out the laryngotracheal region segmentation of an obtained CT image, and generates a laryngotracheal segmentation mask; the laryngotracheal reconstruction module is configured to extract a laryngotracheal center line by adopting a method of combining morphological processing and curve optimization based on the segmentation mask and generate a three-dimensional straightened image; the feature extraction module is configured to extract depth features and geometric features from the straightened image and fuse the depth features and the geometric features; and the stenosis degree evaluation module is configured to predict the stenosis degree of the laryngeal trachea through a deep learning network based on the fusion features to obtain a detection result. According to the system disclosed by the invention, the laryngotracheal region can be accurately extracted, and the stenosis degree of each layer is accurately calculated in combination with an airway straightening method, so that automatic and accurate stenosis grading is provided.

Description

Technical Field

[0001] The present disclosure relates to the technical field of automated image analysis and processing, and more specifically, to a laryngotracheal stenosis detection system and method. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] Laryngotracheal stenosis (LTS) refers to abnormal stenosis of the larynx (including the glottis and subglottic region) and the trachea, resulting in airflow restriction, causing dyspnea, hoarseness, and even severely affecting the ventilation function. The detection of laryngotracheal stenosis mainly relies on CT images and fiberoptic bronchoscopes, and doctors judge the degree of stenosis based on subjective experience. However, this method has problems such as low efficiency and poor result consistency, and it is difficult to meet the rapid and high-precision clinical needs. With the continuous development of high-resolution CT and deep learning technologies, automated image analysis has gradually been applied in the medical field. In existing research, deep learning models (such as 3D U-Net) have been used for airway segmentation, but most of them focus on the overall analysis of the pulmonary airway, and there is a lack of research on the local precise extraction and stenosis grading of the laryngotrachea. In addition, although existing two-dimensional cross-sectional analysis and computational fluid dynamics (CFD) technologies can assist in evaluating the degree of stenosis, they have defects such as single feature extraction, complex calculation, and long time consumption, and cannot meet the needs of real-time diagnosis in clinical practice.

[0004] With the continuous progress of imaging and artificial intelligence technologies, an automated laryngotracheal stenosis grading system has become an important direction for future medical image analysis. The inventors found in their research that the prior art still cannot effectively solve many problems in laryngotracheal segmentation and stenosis grading: 1) Existing airway segmentation methods mostly rely on morphological processing or traditional deep learning technologies. When these methods extract the laryngotracheal region, they are easily interfered by noise and complex anatomical structures. Especially when there are large anatomical differences or the patient's body position is skewed, it is difficult to perform effective automated correction. Traditional deep learning methods usually lack adaptability to changes in the patient's body position, resulting in low accuracy in extracting the airway region, which affects subsequent stenosis grading analysis; although existing segmentation technologies can extract the airway region to a certain extent, they are still affected by noise, anatomical differences, and patient body position skew, resulting in inaccurate extraction of the laryngotracheal region, thereby affecting the accuracy of stenosis grading.

[0005] 2) Most of the existing stenosis grading methods rely on two-dimensional cross-sectional analysis and evaluate through a single two-dimensional cross-sectional feature (such as stenosis area), ignoring the overall characteristics of the airway in three-dimensional space. This method can only reflect the stenosis at a certain position of the airway and cannot comprehensively display the morphological changes of the airway, resulting in insufficient grading accuracy and being unable to effectively evaluate the actual impact of stenosis on respiratory function.

[0006] 3) Although the method based on computational fluid dynamics (CFD) can assist in evaluating the ventilation function of the airway to a certain extent, its calculation is complex, time-consuming, and the feature extraction method is single, resulting in its inapplicability for real-time diagnosis. Although this method can provide a quantitative evaluation of airway stenosis, its calculation timeliness and accuracy are still difficult to meet the needs of clinical real-time diagnosis. Summary of the Invention

[0007] To solve the above problems, the present disclosure proposes a laryngotracheal stenosis detection system and method, which can accurately extract the laryngotracheal region, and combined with the airway straightening method, accurately calculate the stenosis degree of each layer, so as to provide an automated and accurate stenosis grading, solve the problems of insufficient stenosis evaluation accuracy and low automation degree in the prior art, and significantly improve the diagnostic efficiency and grading accuracy.

[0008] To achieve the above object, the present disclosure adopts the following technical solutions: One or more embodiments provide an automatic laryngotracheal stenosis evaluation system, including: A laryngotracheal segmentation module configured to segment the laryngotracheal region of the acquired CT image to generate a laryngotracheal segmentation mask; A laryngotracheal reconstruction module configured to extract the laryngotracheal centerline based on the segmentation mask by combining morphological processing and curve optimization, and generate a three-dimensional straightened image; A feature extraction module configured to extract depth features and geometric features from the straightened image and fuse them; A stenosis degree evaluation module configured to predict the laryngotracheal stenosis degree based on the fused features through a deep learning network to obtain a detection result.

[0009] One or more embodiments provide an automatic laryngotracheal stenosis evaluation method, including the following steps: Segment the laryngotracheal region of the acquired CT image to generate a laryngotracheal segmentation mask; Based on the segmentation mask, extract the laryngotracheal centerline by combining morphological processing and curve optimization, and generate a three-dimensional straightened image; Extract depth features and geometric features from the straightened image and fuse them; Based on the fused features, predict the laryngotracheal stenosis degree through a deep learning network to obtain a detection result.

[0010] Compared with the prior art, the beneficial effects of the present disclosure are as follows: In the present disclosure, through the segmentation of the laryngotrachea and the reconstruction of the three-dimensional straightened image of the laryngotrachea, the stenosis degree of each layer can be accurately calculated, avoiding the limitation of only relying on two-dimensional cross-section analysis in the prior art, comprehensively evaluating the change of the airway morphology, thereby improving the accuracy of stenosis grading and being able to more accurately reflect the stenosis degree.

[0011] In this embodiment, the laryngotracheal reconstruction module refines the morphological features of the laryngotrachea by accurately extracting the centerline of the laryngotrachea and straightening the image, ensuring that it can be accurate to small stenosis sites, eliminating the geometric distortion caused by the morphological change of the laryngotrachea, thereby providing a standardized data basis for subsequent feature extraction and making the subsequent evaluation more efficient and accurate. By reconstructing the laryngotrachea to form an image with standardized dimensions, the influence of anatomical differences between different patients on subsequent analysis can be reduced, and at the same time, the problem of inaccurate cross-sectional measurement caused by patient deviation, inconsistent scanning parameters, etc. can be solved.

[0012] This application optimizes the image processing and feature extraction process through deep learning and has higher computational efficiency compared with CFD. The system can complete the segmentation, centerline extraction and three-dimensional straightening processing of the laryngotracheal region in a short time, so as to quickly obtain the stenosis evaluation result of the laryngotrachea.

[0013] The advantages of the present disclosure and the advantages of additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the present disclosure. The schematic embodiments and descriptions thereof of the present disclosure are used to explain the present disclosure and do not constitute a limitation to the present disclosure.

[0015] Figure 1 is a block diagram of the automatic laryngotracheal stenosis evaluation system according to Embodiment 1 of the present disclosure; Figure 2 is a schematic flowchart of the stenosis degree evaluation according to Embodiment 2 of the present disclosure; Figure 3 is a schematic structural diagram of the ABF module according to Embodiment 1 of the present disclosure; DETAILED DESCRIPTION OF THE EMBODIMENTS The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0016] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0017] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or their combinations. It should be noted that, without conflict, the various embodiments and features in the present disclosure can be combined with each other. The embodiments will be described in detail below with reference to the accompanying drawings.

[0018] Embodiment 1 In the technical solutions disclosed in one or more embodiments, as Figures 1 to 3 shown, an automatic laryngotracheal stenosis assessment system includes: A laryngotracheal segmentation module configured to segment the laryngotracheal region of the acquired CT image to generate a laryngotracheal segmentation mask; A laryngotracheal reconstruction module configured to extract the laryngotracheal centerline based on the segmentation mask by combining morphological processing and curve optimization, and generate a three-dimensional straightened image; A feature extraction module configured to extract depth features and geometric features from the straightened image and fuse them; A stenosis degree assessment module configured to predict the laryngotracheal stenosis degree based on the fused features through a deep learning network to obtain a detection result; In this embodiment, through the laryngotracheal segmentation and the reconstruction of the three-dimensional straightened image of the laryngotrachea, the stenosis degree of each layer can be accurately calculated, avoiding the limitations of only relying on two-dimensional cross-section analysis in the prior art, comprehensively evaluating the changes in the airway morphology, thereby improving the accuracy of stenosis grading and being able to more accurately reflect the stenosis degree.

[0019] Further, an image data import module is also connected and arranged before the laryngotracheal segmentation module, configured to load the CT image data to be processed and perform preprocessing; Optionally, the preprocessing methods include: modal filtering of data, sequence extraction, filtering of the number of slices and reconstruction slice thickness, image window value processing, etc., to ensure that the quality of the input data is suitable for subsequent analysis and avoid errors in the evaluation results caused by noise or low resolution.

[0020] Modal filtering: Detect the modal information of the input image, and only retain the thin-slice CT (slice thickness below 2 mm) scan data, excluding non-CT modalities to ensure that the data meets the analysis requirements; Sequence normalization: Normalize the number of slices, pixel pitch, and slice thickness of the image. Adjust the slice spacing to 1 mm uniformly to ensure that the image has a consistent spatial resolution for subsequent analysis; Window value adjustment: To enhance the visibility of the laryngotracheal region, the image window width and window level are automatically adjusted. The window width is set to 2000 HU and the window level is 400 HU, making the contrast between soft tissues and air more obvious and improving the reliability of the segmentation result.

[0021] In some embodiments, the laryngotracheal segmentation module is used to automatically segment the laryngotracheal region and generate a laryngotracheal segmentation mask. A specific implementation scheme, the laryngotracheal segmentation module includes: Deep learning segmentation unit: Configured to use a pre-trained 3D segmentation deep learning model, the 3D-Unet network, to segment the laryngotracheal region in the CT image. Segmentation optimization unit: Configured to further optimize the segmentation result through post-processing to obtain the final laryngotracheal segmentation mask (mask). The post-processing method includes morphological filtering, connected component analysis to remove isolated small noise regions, region merging, and noise removal in sequence to ensure that the final segmentation result is clear and accurate. Specifically: Morphological operations to remove isolated noise regions, regions smaller than the set volume threshold, where the threshold is set to 1000 mm 3 ; Connectivity analysis to ensure that the laryngotracheal region is the largest connected component and eliminate false segmentation regions. Boundary smoothing: Use dilation and erosion operations to optimize the segmentation boundary, making the segmentation result more natural and continuous.

[0022] It can be achieved that in the deep learning segmentation unit, the prediction output of the 3D-Unet network model includes three categories: background, laryngotrachea, and other tracheas. By constructing a training dataset, the training images are labeled with categories and trained to obtain the 3D-Unet network model. The 3D-Unet network model is used to divide the acquired images into regions to obtain the segmentation region of the laryngotrachea.

[0023] Among them, the laryngotrachea is located in the neck, starting from the lower edge of the larynx (laryngeal prominence), from below the laryngeal prominence (including the epiglottis and vocal cord regions) to the position of the tracheal bifurcation (i.e., where the main bronchi enter the two lungs), and other tracheas include other tracheal regions related to the airway except the laryngotrachea. Specifically, the pre-trained deep learning model 3DU-Net includes an encoder, a bottleneck layer, and a decoder connected in sequence. The encoder gradually extracts spatial features from low-level to high-level through four stages of convolutional blocks and max-pooling operations, while reducing the resolution. Each convolutional block contains two layers of 3×3×3 convolution, batch normalization, and ReLU activation. The bottleneck layer further extracts global semantic information for the extracted spatial features. The decoder gradually restores the spatial resolution through deconvolution and stitches the features with those of the corresponding stage of the encoder through skip connections to retain detailed information. Finally, the feature map is mapped to a segmentation result through a 1×1×1 convolutional layer to generate a segmentation mask of the same size as the input. For example, both the input and the input are 128×128×128.

[0024] The image processed by the decoder segments the entire image in a sliding window manner. The segmentation categories include three categories: the background of the part that does not belong to the airway region, the laryngotracheal region, and other tracheal regions. By adding other tracheal regions, the false positives of the laryngotrachea are reduced to prevent other tracheas from being added and affecting the calculation of the stenosis degree.

[0025] The above laryngotracheal extraction module performs accurate regional identification through a 3D-Unet network deep learning model, reducing errors caused by human factors. Through accurate segmentation, more accurate basic data can be provided for subsequent analysis, avoiding segmentation deviations caused by anatomical variations or pathological changes.

[0026] In some embodiments, the laryngotracheal reconstruction module extracts the laryngotracheal centerline based on the segmentation mask and generates a three-dimensional straightened image, including the extraction method of the laryngotracheal centerline and the generation method of the three-dimensional straightened image; A further technical solution is to use a method combining morphological processing and curve optimization to extract the laryngotracheal centerline, including the following steps: 1) For the segmented laryngotracheal segmentation mask image, a method combining distance transformation and morphological processing is used to extract the skeleton base points: Specifically, the extraction of the skeleton base points includes: 1.1) Use Euclidean distance to calculate the distance from each foreground (laryngotracheal region) pixel to the nearest background pixel to obtain a distance map; 1.2) Slide a sliding window on the distance map, compare the values of the center point and the neighborhood points. If the center point is a local maximum, it is marked as a skeleton base point; Optionally, the size of the sliding window can be set to 5×5×5; 1.3) Process the skeleton base points with morphological erosion to reduce over-dense skeleton base points; 1.4) Calculate the mean value of the maximum points and set the points greater than the mean value as the final skeleton base points to ensure that the skeleton base points will not be accidentally deleted in subsequent steps; In this embodiment, a distance map is generated by calculating the Euclidean distance, and the skeleton base points are extracted by combining the local maximum method of a sliding window. This method directly extracts local features through geometric characteristics, ensuring the accuracy and sparsity of the skeleton base points. Morphological erosion is used to reduce over-dense points, and the threshold is set by combining the mean of the maximum values to ensure that the skeleton base points are retained during the skeleton generation process.

[0027] 2) The method of iterative erosion, opening operation and smoothed difference extraction is adopted for skeleton extraction, and the segmentation region is shrunk layer by layer to obtain a complete laryngotracheal skeleton; the specific process is as follows: 2.1) Erosion operation: For the segmented mask image marking the skeleton base points, a morphological erosion operation is performed on the foreground image region to remove the peripheral noise of the laryngotrachea; Optionally, the kernel size can be set to 3×3×3; 2.2) Smoothed opening operation: Apply a morphological opening operation to the eroded region, smooth the boundary by eroding first and then dilating, and remove weak connections and noise regions; 2.3) Extract skeleton: Extract the newly added skeleton part currently by calculating the difference between the eroded region and the result of the opening operation, and finally merge the currently extracted skeleton part with the existing skeleton while avoiding repeated extraction; By repeatedly iterating the process of erosion, opening operation and difference extraction of the skeleton in steps 2.1 to 2.3 above, the outer voxel values of the segmentation region are removed layer by layer, and the skeleton structure shrinks layer by layer; Optionally, the maximum number of iterations can be set or the iteration stops when complete erosion is reached.

[0028] 3) Perform connectivity analysis and fracture repair on the obtained skeleton to obtain a connected skeleton, specifically: 3.1) Identify the fracture points in the skeleton through connected component analysis; 3.2) Use the shortest path algorithm to connect the fractured skeleton segments to ensure the connectivity of the skeleton; In the above implementation method, the skeleton extraction process is more refined. Through the local maximum method and morphological erosion, redundant points and noise points in the skeleton generation are avoided; the iterative method of layer-by-layer erosion ensures the multi-level expression of the skeleton and avoids local missing problems caused by one-time skeleton extraction; by repairing the fracture points through connectivity analysis, it is applicable to complex tubular structures and ensures the connectivity and continuity of the skeleton.

[0029] 4) Smooth the centerline of the skeleton, use B-spline fitting to smooth the curve, eliminate sharp changes, and thus obtain a continuous and smooth centerline; This embodiment proposes to use the B-spline fitting method to smooth the extracted centerline, with a focus on the continuity of the curve and the elimination of sharp changes. The smoothing process is globally applied to the extracted centerline rather than local path optimization to ensure the overall smoothness of the extracted centerline.

[0030] A further technical solution is a method for generating a three-dimensional straightened image. Based on the extracted laryngeal tracheal centerline, its tortuous or irregular shape is unfolded into a standardized straight-line structure to eliminate morphological distortion and provide a consistent analysis basis for subsequent feature extraction. The method includes the following steps: 5) For the extracted laryngeal tracheal centerline, sample it and construct a cross-sectional coordinate system, and construct a two-dimensional grid with a set resolution on each cross-section: 5.1) Uniformly sample the centerline at a set sampling interval; Specifically, the sampling interval can be set to the millimeter level, such as 1 mm; the sampling points obtained after equidistant sampling are arranged at equal intervals; 5.2) Construct a cross-sectional plane from the sampling points of the centerline. Each plane defines the sampling points on the centerline as the plane center, and the tangent vector represents the plane normal. Generate a two-dimensional coordinate system of the cross-section through the orthogonal basis of the normal vector; and construct a two-dimensional grid with a set resolution on each cross-section for subsequent image mapping; Optionally, the set resolution of the two-dimensional grid can be 60×60; 6) Based on the obtained cross-sectional coordinate system, perform mapping to intercept the two-dimensional slice image corresponding to each cross-section from the original CT image: 6.1) Transform the grid point coordinates from the local cross-sectional plane to the global three-dimensional coordinate system, and calculate the position of each two-dimensional grid in the global three-dimensional coordinate system to find the corresponding voxel value in the original CT image; 6.2) Use the trilinear interpolation method to resample (1 mm) the voxel values of the original image and map them onto each cross-sectional grid, thereby obtaining the two-dimensional slice image of the plane; that is, use the trilinear interpolation method to extract the corresponding voxel values from the original CT image according to the transformed coordinates and fill them into the two-dimensional grid of the corresponding cross-section.

[0031] 7) Three-dimensional straightened image reconstruction: Stack the two-dimensional slice images corresponding to all cross-sections layer by layer along the centerline direction according to the centerline direction to form a three-dimensional straightened image; The size of the three-dimensional straightened image is L×60×60, where L is the number of sampling points on the centerline, and 60×60 is the resolution of each cross-section.

[0032] In the above embodiment, the fixed-resolution grid and the interpolation resampling method provide high-quality three-dimensional images, laying a solid foundation for subsequent analysis.

[0033] In this embodiment, the laryngotracheal reconstruction module refines the morphological features of the laryngotrachea by accurately extracting the centerline of the laryngotrachea and straightening the image, ensuring that it can be accurate to tiny stenosis sites, eliminating the geometric distortion caused by the morphological changes of the laryngotrachea, thus providing a standardized data basis for subsequent feature extraction and making the subsequent evaluation more efficient and accurate. By reconstructing the laryngotrachea to form an image with standardized dimensions, the influence of anatomical differences between different patients on subsequent analysis can be reduced, and at the same time, the problem of inaccurate cross-sectional measurement caused by patient skew, inconsistent scanning parameters, etc. can be solved.

[0034] In some embodiments, in the feature extraction module, for the obtained three-dimensional straightened image, geometric features are extracted, and the specific process is as follows: 8.1) Extract the major axis, minor axis, and lumen area of the laryngotrachea in each two-dimensional slice image along the centerline of the three-dimensional straightened image, and perform normalization processing; According to the shape of each cross-section, calculate the lengths of the longest axis and the shortest axis respectively to quantify the morphological characteristics of the cross-section as the major axis and the minor axis; use the image integration method to calculate the total area of all pixels within the cross-section to obtain the actual lumen area of each cross-section.

[0035] Among them, during the normalization process, the maximum values max of the major axis, minor axis, and lumen area can be fixed at 50, 50, and 2000 respectively; 8.2) Connect the normalized geometric features through a fully connected layer to obtain a feature vector of a set size, and use this vector as the geometric feature (geo feature) of the laryngotracheal plane of this layer; Specifically, taking the set size of the feature vector as 16 for example, the input of the fully connected layer is 3 and the output is 16, so as to obtain a feature vector of size 16.

[0036] In the feature extraction module, for the obtained three-dimensional straightened image, depth features are extracted. Specifically: sample a sampling block of a set size along the centerline direction from the straightened image, and extract depth features (cnn feature) through a 3D CNN model to capture the spatial structure information of the laryngotrachea; for example, the size of the sampling block is set to 30×30×30; Furthermore, in the feature extraction module, the extracted geometric features and depth features can be fused, which can be implemented by an ABF module, such as Figure 3As shown, first, the geometric features are standardized (Norm) and concatenated into a vector. The ABF module concatenates the geometric features and depth features into a long vector, and then calculates the attention weights of each feature through a two-layer fully connected layer (144→64→144). The weighting coefficient of each feature is obtained through the Sigmoid activation function, and finally, the weighted features are obtained through element-wise weighting. This weighted feature can be used as the input of the subsequent network, enabling the network to dynamically adjust their contributions according to the importance of different features.

[0037] In this embodiment, the depth feature and the geometric feature are fused into an input feature vector. The fused feature is more representative and can provide a more comprehensive information input, effectively improving the accuracy of subsequent evaluation.

[0038] The feature extraction module of this embodiment extracts the depth feature and geometric feature from the straightened image and fuses them. This module extracts the spatial depth feature in the reconstructed image to analyze the overall morphology of the laryngotrachea, and comprehensively describes the shape and size of the laryngotrachea by calculating geometric features such as the major axis, minor axis, and lumen area. The depth feature provides the hierarchical information of the laryngotracheal tissue, and the geometric feature quantifies the physical morphology of the laryngotrachea from a geometric perspective.

[0039] In some embodiments, the stenosis degree evaluation module can input the fused feature into the Transformer network to predict the stenosis degree of each layer of the laryngotrachea and generate a stenosis distribution map; The Transformer network uses the multi-head attention mechanism to capture long-range dependencies, so as to better analyze the subtle changes of the laryngotrachea and predict the stenosis degree. In this embodiment, the input of the Transformer module is B×30×144, where B is the batch size, 144 is the length of the feature vector output by the ABF module, and 30 is the length of the input sequence; Furthermore, during the training of the Transformer, random sampling is performed on the input sequence. The random sampling method is as follows: randomly sample the starting point s, the random range of the starting point is 0~L-30, and the random interval is 1~(L-s) / 30. The input sequence of the Transformer is augmented through these two random sampling methods, thereby improving the generalization ability of the model.

[0040] The output layer of the Transformer network is logistic regression, which directly outputs the stenosis degree of each layer. The stenosis distribution map generated in this step clearly shows the distribution of stenosis in each layer, which helps clinicians to analyze accurately.

[0041] The stenosis evaluation module of this embodiment predicts the degree of laryngotracheal stenosis based on the fused features through a Transformer network. The Transformer network deeply models the fused features at multiple levels through the self-attention mechanism, so that the stenosis conditions in various regions of the laryngotrachea can be accurately captured, avoiding the limitations of traditional methods. By using the multi-head attention mechanism, the system can analyze the morphological changes of the laryngotrachea from different perspectives and improve the evaluation accuracy.

[0042] Furthermore, it also includes a diagnostic report generation module: generating a diagnostic report according to the evaluation results, including information such as a stenosis distribution map and global stenosis information, and can also be classified according to the stenosis rate, such as divided into four levels to obtain the Myer-cotton f classification: Grade I (mild stenosis, stenosis rate < 50%); Grade II (moderate stenosis, 50% - 70%); Grade III (severe stenosis, 71% - 99%); Grade IV (extremely severe stenosis, 100%).

[0043] An automatic laryngotracheal stenosis evaluation system provided by this embodiment realizes accurate segmentation, vascular reconstruction and quantitative evaluation of the stenosis degree in the laryngotracheal region, reduces human intervention, and improves the stenosis detection efficiency and accuracy. Through this system and method, the stenosis site of the laryngotrachea can be quickly and accurately identified in clinical images, and strong data support can be provided for subsequent clinical treatment. By using a deep learning model and automated evaluation technology, the present invention can achieve rapid and efficient stenosis grading, avoiding the complex and time-consuming problems of the computational fluid dynamics method.

[0044] Embodiment 2 Based on Embodiment 1, an automatic laryngotracheal stenosis evaluation method is provided in this embodiment, which is characterized by including the following steps: Step 1: Segment the laryngotracheal region of the acquired CT image to generate a laryngotracheal segmentation mask; Step 2: Based on the segmentation mask, use a method combining morphological processing and curve optimization to extract the laryngotracheal centerline and generate a three-dimensional straightened image; Step 3: Extract depth features and geometric features from the straightened image and fuse them; Step 4: Based on the fused features, predict the degree of laryngotracheal stenosis through a deep learning network to obtain the detection result.

[0045] Furthermore, in Step 1, the method for generating the laryngotracheal segmentation mask includes the following steps: Step 11: Use a pre-trained 3D segmentation deep learning model, the 3D-Unet network, to segment the laryngotracheal region in the CT image; Step 12: Further optimize the segmentation result through post - processing to obtain the final laryngotracheal segmentation mask; the post - processing method includes performing morphological filtering, connected component analysis to remove isolated small noise regions, region merging, and noise removal in sequence.

[0046] In step 2, the method of extracting the laryngotracheal centerline by combining morphological processing and curve optimization includes: Step 21: For the segmented laryngotracheal segmentation mask image, extract the skeleton base points by combining distance transformation and morphological processing: Step 22: Use the methods of iterative erosion, opening operation, and smoothed difference extraction to perform skeleton extraction, gradually shrinking the segmentation region layer by layer to obtain the complete laryngotracheal skeleton; Step 23: Perform connectivity analysis and fracture repair on the obtained skeleton to obtain a connected skeleton; Step 24: Smooth the skeleton centerline, use B - spline to fit the smooth curve, and eliminate sharp changes to obtain a centerline with continuous and smooth lines.

[0047] In step 2, the method for generating a three - dimensional straightened image includes the following steps: Step 25: For the extracted laryngotracheal centerline, perform sampling and construct a cross - sectional coordinate system, and construct a two - dimensional grid with a set resolution on each cross - section; Step 26: Based on the obtained cross - sectional coordinate system, perform mapping, and intercept the two - dimensional slice image corresponding to each cross - section from the original CT image; Step 27: Stack the two - dimensional slice images corresponding to all cross - sections layer by layer along the centerline direction according to the direction of the centerline to form a three - dimensional straightened image.

[0048] It should be noted here that each step in this embodiment corresponds one by one to each module in Embodiment 1, and the specific implementation process is the same, so it will not be repeated here.

[0049] The above are only the preferred embodiments of the present disclosure and are not used to limit the present disclosure. For those skilled in the art, the present disclosure can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

[0050] Although the specific implementation manners of the present disclosure are described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that based on the technical solutions of the present disclosure, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present disclosure.

Claims

1. An automatic evaluation system for laryngeal tracheal stenosis, characterized in that: include: A laryngotracheal segmentation module is configured to segment the laryngotracheal region of the acquired CT image and generate a laryngotracheal segmentation mask; The laryngotracheal reconstruction module is configured to extract the laryngotracheal centerline based on the segmentation mask by combining morphological processing and curve optimization, and generate a three-dimensional straightened image; A feature extraction module is configured to extract depth features and geometric features from the straightened image and fuse them; The stenosis degree assessment module is configured to predict the degree of laryngeal tracheal stenosis based on the fusion features through a deep learning network to obtain a detection result.

2. The automatic evaluation system for laryngotracheal stenosis according to claim 1, characterized in that: Laryngotracheal segmentation module, including: Deep learning segmentation unit: configured to segment the laryngotracheal region in the CT image using a pre-trained 3D segmentation deep learning model 3D-Unet network; Segmentation optimization unit: configured to further optimize the segmentation result through post-processing to obtain the final larynx and trachea segmentation mask; the post-processing method includes sequentially performing morphological filtering, connected domain analysis to remove isolated small noise areas, area merging and noise removal.

3. The automatic evaluation system for laryngotracheal stenosis according to claim 1, characterized in that: The method of extracting the laryngotracheal centerline by combining morphological processing and curve optimization includes the following steps: For the segmented larynx and trachea segmentation mask image, a method combining distance transformation and morphological processing is used to extract skeleton base points: The skeleton is extracted by iterative erosion, opening operation and smooth difference extraction, and the segmented area is shrunk layer by layer to obtain a complete laryngeal and tracheal skeleton. Performing connectivity analysis and fracture repair on the obtained skeleton to obtain a connected skeleton; Smooth the center line of the skeleton, use B-spline to fit the smooth curve, eliminate sharp changes, and obtain a center line with continuous and smooth lines.

4. The automatic evaluation system for laryngotracheal stenosis according to claim 3, characterized in that: The skeleton base point extraction method includes: Use Euclidean distance to calculate the distance from each foreground pixel to the nearest background pixel to obtain a distance map; Use a sliding window to slide on the distance map, compare the values ​​of the center point and the neighborhood points, and if the center point is a local maximum, it is marked as a skeleton base point; Use morphological corrosion to process skeleton base points to reduce over-dense skeleton base points; Calculate the mean of the maximum points and set the points greater than the mean as the final skeleton base points to ensure that the skeleton base points will not be mistakenly deleted in subsequent steps; Alternatively, the skeleton is extracted by iterative corrosion, opening operation and smooth difference extraction, and the segmented area is shrunk layer by layer to obtain a complete laryngeal tracheal skeleton. The specific process is as follows: Step 2.1, for the segmentation mask image with the skeleton base points marked, a morphological corrosion operation is performed on the foreground image area to remove the peripheral noise of the larynx and trachea; Step 2.2: Apply morphological opening operation to the eroded area, smooth the boundary by first corroding and then dilating, and remove weak connections and noise areas; Step 2.3, extract the newly added skeleton part by calculating the difference between the erosion area and the opening operation result, and finally merge the currently extracted skeleton part with the existing skeleton; Repeat the iterative process of erosion, opening and difference skeleton extraction from step 2.1 to step 2.3 above, remove the outer voxel values ​​of the segmented area layer by layer, and shrink the skeleton structure layer by layer.

5. The automatic evaluation system for laryngotracheal stenosis according to claim 1, characterized in that: The method for generating a three-dimensional straightened image comprises the following steps: For the extracted laryngotracheal centerline, sampling is performed and a cross-sectional coordinate system is constructed, and a two-dimensional grid with a set resolution is constructed on each cross section; Mapping is performed based on the obtained cross-sectional coordinate system, and a two-dimensional slice image corresponding to each cross section is intercepted from the original CT image; The two-dimensional slice images corresponding to all cross sections are stacked layer by layer along the center line direction to form a three-dimensional straightened image.

6. The automatic evaluation system for laryngotracheal stenosis according to claim 5, characterized in that: The method for sampling and constructing a cross-sectional coordinate system for the extracted laryngotracheal centerline, and constructing a two-dimensional grid with a set resolution on each cross section, includes: The center line is sampled evenly at the set sampling interval; Construct cross-sectional planes from sampling points on the center line. Each plane defines the sampling points on the center line as the plane center. The tangent vector represents the plane normal. The two-dimensional coordinate system of the cross section is generated through the orthogonal basis of the normal vector, and a two-dimensional grid of a set resolution is constructed on each cross section. Alternatively, mapping is performed based on the obtained cross-sectional coordinate system, and a two-dimensional slice image corresponding to each cross section is intercepted from the original CT image, including: Transform the grid point coordinates from the local cross-sectional plane to the global three-dimensional coordinate system, and calculate the position of each two-dimensional grid in the global three-dimensional coordinate system in order to find the corresponding voxel value in the original CT image; Using the linear interpolation method, the corresponding voxel values ​​are extracted from the original CT image according to the transformed coordinates and filled into the two-dimensional grid of the corresponding cross section.

7. The automatic evaluation system for laryngotracheal stenosis according to claim 1, characterized in that: In the feature extraction module, geometric features are extracted from the obtained 3D straightened image. The specific process is as follows: The long diameter, short diameter and lumen area of ​​the laryngotrachea in each two-dimensional slice image along the center line of the three-dimensional straightened image are extracted and normalized; The normalized geometric features are connected through a fully connected layer to obtain a feature vector of a set size, and the vector is used as the geometric feature of the laryngeal tracheal plane of the layer; Alternatively, in the feature extraction module, the depth features are extracted by using a 3D CNN model for the obtained three-dimensional straightened image; The extracted geometric features and depth features are fused using the ABF module.

8. A method for automatically assessing laryngotracheal stenosis, characterized in that: The steps include: Perform laryngotracheal region segmentation on the acquired CT image to generate a laryngotracheal segmentation mask; Based on the segmentation mask, the morphological processing and curve optimization method are combined to extract the centerline of the larynx and trachea, and generate a three-dimensional straightened image. Extract deep features and geometric features from the straightened image and fuse them; Based on the fusion features, the degree of laryngeal tracheal stenosis is predicted through a deep learning network to obtain the detection results.

9. The method for automatically assessing laryngotracheal stenosis according to claim 8, characterized in that: The method of extracting the centerline of the larynx and trachea by combining morphological processing and curve optimization includes: For the segmented larynx and trachea segmentation mask image, a method combining distance transformation and morphological processing is used to extract skeleton base points: The skeleton is extracted by iterative erosion, opening operation and smooth difference extraction, and the segmented area is shrunk layer by layer to obtain a complete laryngeal and tracheal skeleton. Performing connectivity analysis and fracture repair on the obtained skeleton to obtain a connected skeleton; Smooth the center line of the skeleton, use B-spline to fit the smooth curve, eliminate sharp changes, and obtain a center line with continuous and smooth lines.

10. The method for automatically assessing laryngotracheal stenosis according to claim 8, characterized in that: The method for generating a three-dimensional straightened image comprises the following steps: For the extracted laryngotracheal centerline, sampling is performed and a cross-sectional coordinate system is constructed, and a two-dimensional grid with a set resolution is constructed on each cross section; Mapping is performed based on the obtained cross-sectional coordinate system, and a two-dimensional slice image corresponding to each cross section is intercepted from the original CT image; The two-dimensional slice images corresponding to all cross sections are stacked layer by layer along the center line direction to form a three-dimensional straightened image.

Images