An otolaryngology CT examination image segmentation method and system

Abstract

The embodiment of the present application relates to the technical field of image segmentation, and specifically discloses an otolaryngology CT examination image segmentation method and system. The embodiment of the present application collects public CT images of otolaryngology based on big data technology, and performs privacy processing, expansion and labeling processing; a basic segmentation architecture is selected to construct an image segmentation model; the target CT image is subjected to image preprocessing and initial segmentation processing to generate a plurality of initial segmentation images; the initial segmentation images are imported into the image segmentation model, a plurality of model segmentation images are exported, and fine adjustment is performed to generate a plurality of fine segmentation images. The image segmentation model can be constructed, the target CT image can be subjected to image preprocessing and initial segmentation processing, and then the image segmentation model is used for image segmentation and fine adjustment to generate a plurality of fine segmentation images, so that automatic and accurate otolaryngology CT examination image segmentation can be realized, the cost is reduced, the time consumption is short, and the real-time demand in a clinical environment can be met.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, and in particular relates to a method and system for segmenting CT images in otolaryngology. Background Technology

[0002] Otolaryngology CT image segmentation is an important field in medical image processing. Its goal is to segment different anatomical structures (such as the nasal cavity, larynx, cochlea, etc.) in otolaryngology CT images using algorithms or artificial intelligence techniques, in order to support clinical diagnosis and treatment.

[0003] In the current technology, there is no precise segmentation technology for CT examination images in otolaryngology. Often, only coarse segmentation of CT examination images can be performed. Then, the image segmentation needs to be optimized and refined manually, which greatly increases the manual cost of segmenting CT examination images in otolaryngology and is time-consuming, and cannot meet the real-time needs in clinical environments. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for segmenting CT images in otolaryngology, aiming to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A method for segmenting CT images in otolaryngology, the method specifically includes the following steps:

[0007] Based on big data technology, multiple publicly available CT images from the ENT department are collected, and these publicly available CT images are processed for privacy, expansion, and annotation to generate multiple annotated CT images.

[0008] A basic segmentation architecture is selected, and the basic segmentation architecture is trained and validated based on multiple labeled CT images to construct an image segmentation model;

[0009] Acquire a target CT image for otolaryngology, perform image preprocessing and initial segmentation on the target CT image, and generate multiple initial segmentation images;

[0010] Multiple initial segmented images are imported into the image segmentation model, multiple model segmented images are exported, and fine adjustments are made to generate multiple fine segmented images.

[0011] An image segmentation system for ENT CT examinations, the system comprising a disclosed image processing unit, a model training and validation unit, an image preprocessing and segmentation unit, and a segmentation fine adjustment unit, wherein:

[0012] The publicly available image processing unit is used to collect multiple publicly available CT images of otolaryngology based on big data technology, and to perform privacy, expansion and annotation processing on the multiple publicly available CT images to generate multiple annotated CT images;

[0013] The model training and validation unit is used to select a basic segmentation architecture, train and validate the basic segmentation architecture based on multiple labeled CT images, and construct an image segmentation model.

[0014] The image preprocessing and segmentation unit is used to acquire the target CT image of the ENT department, perform image preprocessing and initial segmentation on the target CT image, and generate multiple initial segmented images.

[0015] The segmentation fine adjustment unit is used to import multiple initial segmented images into the image segmentation model, export multiple model segmented images, and perform fine adjustments to generate multiple fine segmented images.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] This invention utilizes big data technology to collect publicly available CT images from ENT (Ear, Nose, and Throat) departments, performing privacy protection, expansion, and annotation. A basic segmentation architecture is selected to construct an image segmentation model. The target CT image undergoes preprocessing and initial segmentation to generate multiple initial segmented images. These initial segmented images are then imported into the image segmentation model, exported, and finely adjusted to generate multiple finely segmented images. This method enables automatic and accurate segmentation of ENT CT examination images, reducing costs and time consumption, and meeting real-time needs in clinical settings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0019] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0020] Figure 2 A flowchart illustrating the method for generating multiple annotated CT images provided in an embodiment of the present invention is shown.

[0021] Figure 3 A flowchart illustrating the construction of an image segmentation model in the method provided by an embodiment of the present invention is shown.

[0022] Figure 4 A flowchart of image preprocessing and initial segmentation processing in the method provided by an embodiment of the present invention is shown.

[0023] Figure 5 A flowchart illustrating the initial segmentation process in the method provided by an embodiment of the present invention is shown.

[0024] Figure 6 A flowchart illustrating the generation of multiple finely segmented images in the method provided by an embodiment of the present invention is shown.

[0025] Figure 7 An application architecture diagram of the system provided in an embodiment of the present invention is shown.

[0026] Figure 8 A structural block diagram of the image processing unit disclosed in the system provided by an embodiment of the present invention is shown.

[0027] Figure 9 A structural block diagram of the image preprocessing segmentation unit in the system provided by an embodiment of the present invention is shown.

[0028] Figure 10 A structural block diagram of the segmentation fine adjustment unit in the system provided by an embodiment of the present invention is shown. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0030] Understandably, in the current technology, there is no precise segmentation technology for ENT CT examination images. Often, only coarse segmentation of CT examination images can be performed. Then, the image segmentation needs to be optimized and refined manually, which greatly increases the manual cost of ENT CT examination image segmentation and is time-consuming, and cannot meet the real-time needs in the clinical environment.

[0031] To address the aforementioned issues, this invention utilizes big data technology to collect multiple publicly available CT images from the ENT field. These publicly available CT images undergo privacy enhancement, expansion, and annotation processing to generate multiple annotated CT images. A basic segmentation architecture is selected and trained and validated based on the annotated CT images to construct an image segmentation model. A target ENT CT image is acquired, and preprocessed and initially segmented to generate multiple initial segmented images. These initial segmented images are then imported into the image segmentation model, exporting multiple model-segmented images, which are then finely adjusted to generate multiple finely segmented images. This method enables automatic and accurate segmentation of ENT CT examination images, reducing costs and time consumption, and meeting real-time requirements in clinical settings.

[0032] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0033] Specifically, in one embodiment of the present invention, a method for segmenting CT images in otolaryngology includes the following steps:

[0034] Step S101: Based on big data technology, collect multiple publicly available CT images from the Department of Otolaryngology, and perform privacy, expansion and annotation processing on the multiple publicly available CT images to generate multiple annotated CT images.

[0035] In this embodiment of the invention, based on big data technology, multiple publicly available medical databases and / or hospital clinical databases are identified as multiple published data sources. Then, multiple publicly available CT images of otolaryngology are collected from these multiple published data sources. By performing privacy processing on these multiple publicly available CT images to remove private data such as names and hospitals, multiple private CT images are generated. Then, geometric enhancement and intensity enhancement expansion processing is performed on these multiple private CT images through rotation, translation, scaling, flipping, brightness adjustment, and Gaussian noise adjustment to generate multiple expanded CT images. Finally, the segmentation annotations of these multiple expanded CT images are received to generate multiple annotated CT images.

[0036] Specifically, Figure 2 A flowchart illustrating the method for generating multiple annotated CT images provided in an embodiment of the present invention is shown.

[0037] In a preferred embodiment of the present invention, the step of collecting multiple publicly available CT images from the ENT department based on big data technology, and performing privacy, expansion, and annotation processing on the multiple publicly available CT images to generate multiple annotated CT images specifically includes the following steps:

[0038] Step S1011: Based on big data technology, identify multiple publishing data sources;

[0039] Step S1012: Collect multiple publicly available CT images of otolaryngology from multiple published data sources;

[0040] Step S1013: Perform privacy processing on the multiple publicly available CT images to generate multiple privacy CT images;

[0041] Step S1014: Perform geometric enhancement and intensity enhancement expansion processing on the multiple privacy CT images to generate multiple expanded CT images;

[0042] Step S1015: Receive segmentation annotations for the multiple extended CT images and generate multiple annotated CT images.

[0043] In a preferred embodiment of the present invention, the step of performing geometric enhancement and intensity enhancement expansion processing on multiple privacy CT images to generate multiple expanded CT images specifically includes the following steps:

[0044] Based on the position coordinates of each pixel in the privacy CT image, a two-dimensional rotation transformation is performed on the privacy CT image to obtain the transformed privacy CT image.

[0045] Gaussian smoothing was applied to the transformed privacy CT image for denoising, resulting in a denoised privacy CT image.

[0046] Two-dimensional fast Fourier transform is performed on the denoised privacy CT image to achieve frequency domain enhancement, resulting in a geometrically enhanced privacy CT image.

[0047] The grayscale value of each pixel in the privacy CT image is nonlinearly transformed using gamma correction, and then the nonlinear transformation result is adjusted using exponential decay to obtain the enhanced privacy CT image.

[0048] The enhanced privacy CT image was decomposed into high-frequency and low-frequency components using wavelet decomposition.

[0049] Gaussian noise is added to the sensitive region corresponding to the high-frequency component to protect privacy, resulting in the blurred high-frequency component.

[0050] The low-frequency component and the blurred high-frequency component are subjected to inverse wavelet transform to obtain the enhanced privacy CT image.

[0051] The geometrically enhanced privacy CT image and the intensity-enhanced privacy CT image are weighted and fused, and the fusion result is nonlinearly adjusted to obtain the extended CT image.

[0052] In this embodiment of the invention, a combination of geometric enhancement and intensity enhancement is employed. This not only optimizes the geometric properties of the image (such as rotation, translation, and scaling) but also enhances pixel intensity properties (such as grayscale value and contrast). This two-way approach comprehensively improves image quality, enabling the model to be robust to both spatial and intensity variations. Furthermore, the geometric enhancement process incorporates frequency domain processing (such as Fourier transform), effectively extracting implicit high-frequency features (such as edges and textures) and low-frequency features (such as overall shape) from the image. Compared to processing only in the spatial domain, this method allows for deeper refinement of image features, improving the model's ability to capture important features.

[0053] Frequency domain processing can more efficiently enhance specific parts (such as edges), while the spatial domain is responsible for maintaining the overall visualization quality.

[0054] Furthermore, to ensure both intensity enhancement and privacy protection, wavelet transform was used for multi-scale feature decomposition, bringing more levels of detail to the image. Gaussian blur was also applied to sensitive regions corresponding to high-frequency components to protect privacy. This multi-scale feature extraction method is well-suited for medical image processing, as anatomical structures typically require multi-level representation while maintaining patient privacy.

[0055] In a preferred embodiment of the present invention, the step of receiving segmentation annotations for multiple extended CT images and generating multiple annotated CT images specifically includes the following steps:

[0056] Select one CT image from multiple extended CT images as the current image;

[0057] The pre-trained U-Net is used to segment the current image to obtain a segmentation probability map;

[0058] By introducing manual annotation, and calculating the confidence intervals of different annotation schemes based on the results of multiple annotators, optimized manual annotations are generated.

[0059] The optimized manual annotations are converted into discrete region segmentation images. The confidence scores are dynamically calculated based on whether the discrete region segmentation images are boundary regions. The confidence scores are then converted into weights to obtain the manual confidence weights.

[0060] Assign corresponding confidence levels based on the probability values ​​of the segmentation probability map to obtain the intelligent segmentation confidence level, and convert the intelligent segmentation confidence level into weights to obtain the model confidence weights;

[0061] The segmentation probability map is weighted with the model confidence weight to obtain the first weighted result. The manual annotation is weighted with the manual confidence weight to obtain the second weighted result. The current image is fused with the corresponding first weighted result and second weighted result to obtain the fused annotated image.

[0062] Multi-scale optimization is performed on the fused annotated images to obtain the final annotated CT images.

[0063] In this embodiment of the invention, the segmentation model provides the ability to automatically generate preliminary results, significantly reducing the workload of manual annotation. Human annotation experts only need to correct specific regions, rather than starting from scratch, significantly improving annotation efficiency. Manual annotation in medical images may exhibit some subjective differences (e.g., different annotators have inconsistent judgments of boundaries). By introducing confidence analysis and probabilistic fusion, the system can smooth out the differences between annotators, generating more consistent annotation results. Simultaneously, the results of the segmentation model can also serve as a reference, reducing omissions in manual annotation. Finally, through probabilistic fusion, the system can dynamically adjust the weights based on the confidence levels of intelligent segmentation and manual annotation. Compared to simple methods of "manual annotation covering intelligent segmentation" or "intelligent segmentation replacing manual annotation," probabilistic fusion ensures the flexibility and robustness of the processing results. The dynamic allocation of confidence weights allows the system to make optimal decisions based on the characteristics of different regions (high-confidence regions or low-confidence regions).

[0064] Furthermore, the ENT CT image segmentation method also includes the following steps:

[0065] Step S102: Select a basic segmentation architecture, and train and validate the basic segmentation architecture based on multiple labeled CT images to construct an image segmentation model.

[0066] In this embodiment of the invention, a basic segmentation architecture is selected from multiple network architectures such as U-Net, 3D U-Net, Hybrid CNN-Transformer and Swin UNETR, and the data partitioning ratio is obtained. Then, according to the data partitioning ratio, multiple labeled CT images are divided into a model training set and a model validation set. Based on the model training set and the model validation set, the basic segmentation architecture is trained and validated to construct an image segmentation model.

[0067] Specifically, Figure 3 A flowchart illustrating the construction of an image segmentation model in the method provided by an embodiment of the present invention is shown.

[0068] In a preferred embodiment of the present invention, the step of selecting a basic segmentation architecture, training and validating the basic segmentation architecture based on multiple labeled CT images, and constructing an image segmentation model specifically includes the following steps:

[0069] Step S1021: Select a basic segmentation architecture from a set of preset network architectures;

[0070] Step S1022: Obtain the data partitioning ratio;

[0071] Step S1023: According to the data division ratio, the multiple labeled CT images are divided into a model training set and a model validation set;

[0072] Step S1024: Based on the model training set and the model validation set, train and validate the basic segmentation architecture to construct an image segmentation model.

[0073] In a preferred embodiment of the present invention, the step of training and validating the basic segmentation architecture based on the model training set and the model validation set to construct an image segmentation model specifically includes the following steps:

[0074] The model training set is divided into multiple small batches, each batch containing several training samples;

[0075] Data augmentation is performed on the training samples to obtain augmented images;

[0076] The enhanced image is input into the basic segmentation architecture for forward propagation to perform segmentation prediction and auxiliary task prediction, thereby obtaining the classification result and the target boundary contour.

[0077] The main task uses Dice loss to evaluate the segmentation results, and the auxiliary task uses cross-entropy loss to evaluate the boundary detection results. The Dice loss and cross-entropy loss are weighted and summed, and regularization loss is introduced to obtain a hybrid loss function.

[0078] The hybrid loss function is input into the Adam optimizer for optimization, and the Adam optimizer weight decay strategy and learning parameters are set. The loss is minimized by updating the weights and learning parameters to train the basic segmentation architecture, and the trained basic segmentation architecture is obtained.

[0079] After training, the performance of the trained basic segmentation architecture is verified using a validation set, and the image segmentation model is obtained after verification.

[0080] In a preferred embodiment of the present invention, the step of inputting the enhanced image into the basic segmentation architecture for forward propagation to perform segmentation prediction and auxiliary task prediction, and obtaining the classification result and target boundary contour, specifically includes the following steps:

[0081] Step S102401: The basic segmentation architecture consists of an encoder and a decoder, and both the encoder and decoder are configured with several layers.

[0082] Step S102402: Input the enhanced image into the first layer encoder and extract the local feature map of the enhanced image through the convolutional neural network;

[0083] Step S102403: The multi-scale local detail features are fed into the Transformer module for global modeling to capture long-distance dependencies and obtain a global feature map.

[0084] Step S102404: Perform global average pooling on the local feature map and the global feature map respectively to obtain the average value of each channel;

[0085] Step S102405: Input the average value of each channel into a shared fully connected network to generate coefficients that dynamically adjust the weights of local and global features;

[0086] Step S102406: The coefficients of dynamically adjusted local and global feature weights are compressed to the range of [0, 1] using the Sigmoid activation function to obtain the local feature weights and global feature weights.

[0087] Step S102407: Each channel of the local feature map is multiplied by its corresponding weight to achieve dynamic adjustment of the local features and obtain a weighted local feature map.

[0088] Step S102408: Each channel of the global feature map is multiplied by its corresponding weight to achieve dynamic adjustment of the global features and obtain a weighted global feature map.

[0089] Step S102409: The local feature map and the global feature map are concatenated along the channel dimension, and channel compression is performed through a 1×1 convolution operation to obtain the fused feature map.

[0090] Step S102410: The fused feature map is used as the input to the next layer decoder. Steps S102401 to S102410 are repeated several times in an iterative manner to obtain the final fused features.

[0091] Step S102411: The final fused features are fed into the decoder, and a 3×3 convolution operation is performed on the final fused features to obtain the convolution feature map.

[0092] Step S102412: Perform global average pooling on the convolutional feature map to obtain the global pooling result;

[0093] Step S102413: The global pooling result is passed through a fully connected layer and a Sigmoid activation function to generate channel weights;

[0094] Step S102414: Perform a channel-by-channel weighted operation on the convolutional feature map and channel weights to obtain the output of the first layer decoder;

[0095] In step S102415, dense skip connections are used to concatenate the output of the previous layer decoder with the output of the corresponding layer encoder in the channel dimension as the input of the next layer decoder. Steps S102411 to S102414 are repeated several times in an iterative manner to obtain the final output of the decoder.

[0096] In step S102416, the decoder's final output undergoes an upsampling operation to generate a segmentation map with the same resolution as the input image, resulting in a high-resolution feature map with multi-scale feature aggregation. This high-resolution feature map with multi-scale feature aggregation is then fed into the classification head to generate a class probability distribution and obtain the classification result.

[0097] Step S102417: Take the output of the intermediate layer decoder and perform boundary detection operation to generate a boundary probability image and obtain the target boundary contour.

[0098] In this embodiment of the invention, the advantages of Convolutional Neural Networks (CNN) and Lightweight Transformers are combined, and a feature fusion mechanism is introduced. The CNN portion excels at extracting local information from images, such as edges, textures, and small lesions.

[0099] The Transformer part captures long-distance dependencies through a self-attention mechanism, understanding the global context of the image (such as the overall shape of anatomical structures, left-right symmetry, etc.).

[0100] The fused hybrid features can take into account both the details of tiny targets and the global structural information in pixel-level segmentation tasks, thereby improving the overall segmentation accuracy of the model. In medical imaging, everything from the outline of lesions to the shape and location of the entire anatomical region is very important. This hybrid extraction method can find tiny lesions while maintaining awareness of the global anatomical structure.

[0101] Furthermore, dynamic weight adjustment is introduced during the fusion process, adjusting the relative importance between CNN and Transformer features according to task requirements. For segmentation tasks with blurred boundaries, more reliance may be placed on global Transformer features; for segmentation tasks with rich details, more reliance may be placed on CNN features. This dynamic mechanism makes the model more adaptable, enabling it to achieve better results across different tasks and data distributions. In CT images, lesion boundaries may be blurred or mixed with surrounding tissue, and the contextual information of the anatomical structure can help the model determine the boundary location. This dynamic adjustment ensures that the model can accurately segment boundaries while maintaining global consistency of the context.

[0102] Furthermore, the ENT CT image segmentation method also includes the following steps:

[0103] Step S103: Obtain the target CT image for otolaryngology, perform image preprocessing and initial segmentation on the target CT image, and generate multiple initial segmentation images.

[0104] In this embodiment of the invention, a target CT image for otolaryngology that needs to be segmented is acquired. The target CT image is then subjected to image preprocessing, including noise removal, image enhancement, resolution adjustment, and standardization, to generate a standard CT image. Next, the segmentation candidate regions of the standard CT image are identified, and the segmentation candidate regions in the standard CT image are extracted separately to obtain candidate CT images, thereby eliminating irrelevant image regions. Then, a preset image input size is determined, and the candidate CT image is segmented according to the image input size to obtain multiple initial segmented images.

[0105] Specifically, Figure 4 A flowchart of image preprocessing and initial segmentation processing in the method provided by an embodiment of the present invention is shown.

[0106] In a preferred embodiment of the present invention, the steps of acquiring the target CT image for otolaryngology, performing image preprocessing and initial segmentation on the target CT image, and generating multiple initial segmented images specifically include the following steps:

[0107] Step S1031: Acquire the target CT image for otolaryngology.

[0108] Step S1032: Perform noise removal, image enhancement, resolution adjustment and standardization image preprocessing on the target CT image to generate a standard CT image;

[0109] Step S1033: Perform initial segmentation processing on the standard CT image to generate multiple initial segmented images.

[0110] Specifically, Figure 5 A flowchart illustrating the initial segmentation process in the method provided by an embodiment of the present invention is shown.

[0111] In a preferred embodiment of the present invention, the initial segmentation processing of the standard CT image to generate multiple initial segmented images specifically includes the following steps:

[0112] Step S10331: Identify the segmentation candidate regions of the standard CT image;

[0113] Step S10332: Based on the segmented candidate region, the standard CT image is cropped to obtain a candidate CT image;

[0114] Step S10333: Determine the preset image input size, and perform size segmentation on the candidate CT image to obtain multiple initial segmented images.

[0115] Furthermore, the ENT CT image segmentation method also includes the following steps:

[0116] Step S104: Import the multiple initial segmented images into the image segmentation model, export multiple model segmented images, and perform fine adjustments to generate multiple fine segmented images.

[0117] In this embodiment of the invention, multiple initial segmented images are imported into an image segmentation model, and automatic image segmentation processing is performed by the image segmentation model to export multiple model segmented images. Then, erosion, dilation, opening and closing operations are performed on the multiple model segmented images to repair holes and / or isolated regions, generating multiple repaired segmented images. Finally, the multiple repaired segmented images are subjected to fine edge smoothing processing to generate multiple fine segmented images.

[0118] Specifically, Figure 6 A flowchart illustrating the generation of multiple finely segmented images in the method provided by an embodiment of the present invention is shown.

[0119] In a preferred embodiment of the present invention, the step of importing multiple initial segmented images into the image segmentation model, exporting multiple model segmented images, and performing fine adjustments to generate multiple finely segmented images specifically includes the following steps:

[0120] Step S1041: Import the multiple initial segmented images into the image segmentation model;

[0121] Step S1042: Export multiple model segmentation images;

[0122] Step S1043: Perform erosion, dilation, opening and closing operations on multiple model segmentation images to repair holes and / or isolated regions, generating multiple repaired segmentation images;

[0123] Step S1044: Perform fine edge smoothing processing on the multiple repaired segmented images to generate multiple fine segmented images.

[0124] Furthermore, Figure 7 An application architecture diagram of the system provided in an embodiment of the present invention is shown.

[0125] In another preferred embodiment of the present invention, an ENT CT examination image segmentation system includes:

[0126] The publicly available image processing unit 101 is used to collect multiple publicly available CT images of otolaryngology based on big data technology, and to perform privacy, expansion and annotation processing on the multiple publicly available CT images to generate multiple annotated CT images.

[0127] In this embodiment of the invention, the publicly available image processing unit 101, based on big data technology, identifies multiple publicly available medical databases and / or hospital clinical databases, and then collects multiple publicly available CT images of otolaryngology from these multiple publicly available data sources. By performing privacy processing on the multiple publicly available CT images to remove private data such as names and hospitals, multiple private CT images are generated. Then, through rotation, translation, scaling, flipping, brightness adjustment, Gaussian noise adjustment, and other methods, geometric enhancement and intensity enhancement expansion processing are performed on the multiple private CT images to generate multiple expanded CT images. Finally, the segmentation annotations of the multiple expanded CT images are received to generate multiple annotated CT images.

[0128] Specifically, Figure 8 A structural block diagram of the image processing unit 101 disclosed in the system provided by an embodiment of the present invention is shown.

[0129] In a preferred embodiment provided by the present invention, the disclosed image processing unit 101 specifically includes:

[0130] The data source determination module 1011 is used to determine multiple publishing data sources based on big data technology;

[0131] Image collection module 1012 is used to collect multiple publicly available CT images of otolaryngology from multiple published data sources;

[0132] The privacy processing module 1013 is used to perform privacy processing on multiple publicly available CT images to generate multiple privacy CT images;

[0133] The extended processing module 1014 is used to perform geometric enhancement and intensity enhancement extended processing on multiple privacy CT images to generate multiple extended CT images;

[0134] The segmentation and annotation module 1015 is used to receive segmentation annotations for multiple extended CT images and generate multiple annotated CT images.

[0135] Furthermore, the ENT CT image segmentation system also includes:

[0136] The model training and validation unit 102 is used to select a basic segmentation architecture, train and validate the basic segmentation architecture based on multiple labeled CT images, and construct an image segmentation model.

[0137] In this embodiment of the invention, the model training and validation unit 102 selects a basic segmentation architecture from multiple network architectures such as U-Net, 3D U-Net, HybridCNN-Transformer and Swin UNETR, and obtains the data partitioning ratio. Then, according to the data partitioning ratio, it divides multiple labeled CT images into a model training set and a model validation set. Based on the model training set and the model validation set, it trains and validates the basic segmentation architecture to construct an image segmentation model.

[0138] The image preprocessing and segmentation unit 103 is used to acquire the target CT image of the otolaryngology department, perform image preprocessing and initial segmentation processing on the target CT image, and generate multiple initial segmented images.

[0139] In this embodiment of the invention, the image preprocessing segmentation unit 103 acquires the target CT image of the ENT department that needs to be segmented, performs image preprocessing such as noise removal, image enhancement, resolution adjustment and standardization on the target CT image to generate a standard CT image, identifies the segmentation candidate region of the standard CT image, extracts the segmentation candidate region in the standard CT image separately to obtain candidate CT images, thereby eliminating irrelevant image regions, and then determines the preset image input size. According to the image input size, the candidate CT image is segmented by size to obtain multiple initial segmented images.

[0140] Specifically, Figure 9 A structural block diagram of the image preprocessing segmentation unit 103 in the system provided by an embodiment of the present invention is shown.

[0141] In a preferred embodiment of the present invention, the image preprocessing segmentation unit 103 specifically includes:

[0142] Image acquisition module 1031 is used to acquire target CT images for otolaryngology.

[0143] Image preprocessing module 1032 is used to perform noise removal, image enhancement, resolution adjustment and standardization on the target CT image to generate a standard CT image;

[0144] The initial segmentation module 1033 is used to perform initial segmentation processing on the standard CT image to generate multiple initial segmented images.

[0145] Furthermore, the ENT CT image segmentation system also includes:

[0146] The fine-tuning segmentation unit 104 is used to import multiple initial segmented images into the image segmentation model, export multiple model segmented images, and perform fine-tuning to generate multiple fine-tuned segmented images.

[0147] In this embodiment of the invention, the segmentation fine adjustment unit 104 imports multiple initial segmented images into the image segmentation model, performs automatic image segmentation processing through the image segmentation model, exports multiple model segmented images, and then performs erosion, dilation, opening and closing operations on the multiple model segmented images to repair holes and / or isolated regions, generating multiple repaired segmented images. Subsequently, the multiple repaired segmented images are subjected to fine edge smoothing processing to generate multiple fine segmented images.

[0148] Specifically, Figure 10 A structural block diagram of the segmentation fine adjustment unit 104 in the system provided by an embodiment of the present invention is shown.

[0149] In a preferred embodiment of the present invention, the segmentation fine adjustment unit 104 specifically includes:

[0150] Image import module 1041 is used to import multiple initial segmented images into the image segmentation model;

[0151] Image export module 1042 is used to export multiple model segmentation images;

[0152] Image restoration module 1043 is used to perform erosion, dilation, opening and closing operations on multiple model segmented images to repair holes and / or isolated regions and generate multiple restored segmented images;

[0153] The fine-tuning module 1044 is used to perform fine-tuning processing on the edges of the multiple repaired segmented images to generate multiple fine-tuned segmented images.

[0154] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0155] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0156] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An ear-nose-throat CT examination image segmentation method, characterized by, The method specifically comprises the following steps: Based on big data technology, collect multiple public CT images of otolaryngology, and perform privacy processing, expansion and labeling processing on the multiple public CT images to generate multiple labeled CT images; Select a basic segmentation architecture, train and verify the basic segmentation architecture based on the multiple labeled CT images, and construct an image segmentation model; Obtain a target CT image of otolaryngology, perform image preprocessing and initial segmentation processing on the target CT image, and generate multiple initial segmentation images; Import the multiple initial segmentation images into the image segmentation model, export multiple model segmentation images, and perform fine adjustment to generate multiple fine segmentation images; The method specifically comprises the following steps: Determine multiple publishing data sources based on big data technology; Collect multiple public CT images of otolaryngology from the multiple publishing data sources; Perform privacy processing on the multiple public CT images to generate multiple privacy CT images; Perform geometric enhancement and intensity enhancement expansion processing on the multiple privacy CT images to generate multiple expanded CT images; Receive segmentation labeling for the multiple expanded CT images to generate multiple labeled CT images; The method specifically comprises the following steps: Perform two-dimensional rotation transformation on the privacy CT image according to the position coordinates of each pixel point of the privacy CT image to obtain a transformed privacy CT image; Perform Gaussian smoothing on the transformed privacy CT image to obtain a denoised privacy CT image; Perform two-dimensional fast Fourier transform on the denoised privacy CT image to realize frequency domain enhancement and obtain a geometrically enhanced privacy CT image; Perform nonlinear transformation on the gray value of each pixel point of the privacy CT image using a gamma correction method, and then adjust the nonlinear transformation result using exponential decay to obtain an intensity-enhanced privacy CT image; Wavelet-decompose the intensity-enhanced privacy CT image into high-frequency components and low-frequency components; Add Gaussian noise to the sensitive area corresponding to the high-frequency components to protect privacy and obtain blurred high-frequency components; Perform wavelet inverse transform on the low-frequency components and the blurred high-frequency components to obtain an intensity-enhanced privacy CT image; Weightedly fuse the geometrically enhanced privacy CT image and the intensity-enhanced privacy CT image, and perform nonlinear adjustment on the fusion result to obtain an expanded CT image.

2. The ear-nose-throat CT examination image segmentation method according to claim 1, characterized by, The method specifically comprises the following steps: Select one CT image from the multiple expanded CT images as a current image; Perform segmentation operation on the current image using a pre-trained U-Net to obtain a segmentation probability map; Introduce artificial labeling, calculate the confidence interval of different labeling schemes according to the results of multiple labelers, and generate optimized artificial labeling. The optimized artificial annotation is converted into a discrete region division image, a confidence is dynamically calculated according to whether the discrete region division image is a boundary region, and the confidence is converted into a weight to obtain an artificial confidence weight; According to the probability value of the segmentation probability map, a corresponding confidence is assigned to obtain an intelligent segmentation confidence, and the intelligent segmentation confidence is converted into a weight to obtain a model confidence weight; The segmentation probability map and the model confidence weight are weighted to obtain a first weighted result, the artificial annotation and the artificial confidence weight are weighted to obtain a second weighted result, and the current image and the corresponding first weighted result and the second weighted result are fused to obtain a fused annotation image; The fused annotation image is subjected to multi-scale optimization to obtain a final annotated CT image.

3. The ear-nose-throat CT examination image segmentation method according to claim 2, characterized by, The selecting a basic segmentation architecture, training and verifying the basic segmentation architecture based on the plurality of annotated CT images, and constructing an image segmentation model specifically includes the following steps: selecting a basic segmentation architecture from a plurality of preset network architectures; obtaining a data division ratio; dividing the plurality of annotated CT images into a model training set and a model verification set according to the data division ratio; training and verifying the basic segmentation architecture based on the model training set and the model verification set to construct an image segmentation model.

4. The ear-nose-throat CT examination image segmentation method according to claim 3, characterized by, The training and verifying the basic segmentation architecture based on the model training set and the model verification set to construct an image segmentation model specifically includes the following steps: dividing the model training set into a plurality of small batches, each batch containing a plurality of training samples; performing a data enhancement operation on the training samples to obtain enhanced images; inputting the enhanced images into the basic segmentation architecture for forward propagation to perform segmentation prediction and auxiliary task prediction, and thereby obtaining classification results and target boundary contours; evaluating the segmentation results using a Dice loss for the main task, evaluating the boundary detection results using a cross-entropy loss for the auxiliary task, weighting and summing the Dice loss and the cross-entropy loss, and introducing a regularization loss to obtain a hybrid loss function; inputting the hybrid loss function into an Adam optimizer for optimization, setting an Adam optimizer weight decay strategy and learning parameters; training the basic segmentation architecture by updating the weights and learning parameters to minimize the loss, thereby obtaining a trained basic segmentation architecture; after training is completed, verifying the performance of the trained basic segmentation architecture using the verification set, and obtaining an image segmentation model after verification is completed.

5. The ear-nose-throat CT examination image segmentation method according to claim 4, characterized by, The inputting the enhanced images into the basic segmentation architecture for forward propagation to perform segmentation prediction and auxiliary task prediction, and thereby obtaining classification results and target boundary contours specifically includes the following steps: Step S102401, the basic segmentation architecture is composed of an encoder and a decoder, and the encoder and the decoder each correspondingly have a plurality of layers; Step S102402, inputting the enhanced images into the first layer of the encoder to extract local feature maps of the enhanced images through a convolutional neural network; Step S102403, sending the multi-scale local detail features into a Transformer module for global modeling to capture long-distance dependencies, thereby obtaining global feature maps; Step S102404, global average pooling is performed on the local feature map and the global feature map respectively to obtain the average value of each channel; Step S102405, the average value of each channel is input into a shared fully connected network to generate a coefficient for dynamically adjusting the local feature and the global feature weight; Step S102406, the coefficient for dynamically adjusting the local feature and the global feature weight is compressed to the range of [0, 1] through a Sigmoid activation function to obtain the local feature weight and the global feature weight; Step S102407, each channel of the local feature map is multiplied by the corresponding weight to realize dynamic adjustment of the local feature, and a weighted local feature map is obtained; Step S102408, each channel of the global feature map is multiplied by the corresponding weight to realize dynamic adjustment of the global feature, and a weighted global feature map is obtained; Step S102409, the local feature map and the global feature map are spliced in the channel dimension, and channel compression is performed through a 1×1 convolution operation to obtain a fusion feature map; Step S102410, the fusion feature map is repeatedly input into the next layer encoder in an iterative manner for several times to obtain a final fusion feature; Step S102411, the final fusion feature is input into a decoder, and a 3×3 convolution operation is performed on the final fusion feature to obtain a convolution feature map; Step S102412, global average pooling is performed on the convolution feature map to obtain a global pooling result; Step S102413, the global pooling result is input into a fully connected layer and a Sigmoid activation function to generate a channel weight; Step S102414, the convolution feature map and the channel weight are weighted in each channel to obtain the output of the first layer decoder; Step S102415, the output of the previous layer decoder and the output of the corresponding layer encoder are spliced in the channel dimension through a dense skip connection to serve as the input of the next layer decoder, and steps S102411 to S102414 are repeatedly performed in an iterative manner for several times to obtain the final output of the decoder; Step S102416, the final output of the decoder is upsampled to generate a segmentation map with the same resolution as the input image, a high-resolution feature map with multi-scale feature aggregation is obtained, and the high-resolution feature map with multi-scale feature aggregation is input into a classification head to generate a class probability distribution and obtain a classification result; Step S102417, the output of the intermediate layer decoder is subjected to a boundary detection operation to generate a boundary probability image and obtain a target boundary contour.

6. The ear-nose-throat CT examination image segmentation method according to claim 5, characterized by, The method comprises the following steps: obtaining a target CT image of an ear-nose-throat department; performing image preprocessing and initial segmentation processing on the target CT image to generate a plurality of initial segmentation images, which comprises the following steps: obtaining a target CT image of an ear-nose-throat department; performing image preprocessing of noise removal, image enhancement, resolution adjustment and standardization on the target CT image to generate a standard CT image; performing initial segmentation processing on the standard CT image to generate a plurality of initial segmentation images.

7. The ear-nose-throat CT examination image segmentation method according to claim 6, characterized by, The initial segmentation processing of the standard CT image to generate a plurality of initial segmentation images specifically includes the following steps: Identify the segmentation candidate region of the standard CT image; According to the segmentation candidate region, the standard CT image is intercepted to obtain a candidate CT image; Determine the preset image input size, and size segment the candidate CT image to obtain a plurality of initial segmentation images.

8. The ear-nose-throat CT examination image segmentation method according to claim 7, characterized by, The plurality of initial segmentation images are introduced into the image segmentation model to export a plurality of model segmentation images, and fine adjustment is performed to generate a plurality of fine segmentation images, specifically including the following steps: The plurality of initial segmentation images are introduced into the image segmentation model; Export a plurality of model segmentation images; The plurality of model segmentation images are subjected to erosion, expansion, opening operation and closing operation to repair holes and / or isolated regions to generate a plurality of repaired segmentation images; The plurality of repaired segmentation images are subjected to fine processing of smooth edges to generate a plurality of fine segmentation images.

9. An otolaryngology CT examination image segmentation system, the system being applied to the otolaryngology CT examination image segmentation method according to any one of claims 1 to 8, characterized by The system includes a public image processing unit, a model training and verification unit, an image preprocessing and segmentation unit, and a segmentation fine adjustment unit, wherein: The public image processing unit is used to collect a plurality of public CT images of otolaryngology based on big data technology, and to generate a plurality of labeled CT images by performing privacy, expansion and labeling processing on the plurality of public CT images; The model training and verification unit is used to select a basic segmentation architecture, train and verify the basic segmentation architecture based on the plurality of labeled CT images, and construct an image segmentation model; The image preprocessing and segmentation unit is used to obtain a target CT image of otolaryngology, perform image preprocessing and initial segmentation processing on the target CT image, and generate a plurality of initial segmentation images; The segmentation fine adjustment unit is used to introduce the plurality of initial segmentation images into the image segmentation model to export a plurality of model segmentation images, and perform fine adjustment to generate a plurality of fine segmentation images.

Images