Chronic sinusitis full-automatic segmentation method based on deep learning and related equipment

Abstract

The invention discloses a chronic sinusitis full-automatic segmentation method based on deep learning and related equipment. The method comprises the following steps: acquiring paranasal sinus CT image data of a chronic sinusitis patient, wherein the paranasal sinus CT image data comprises paranasal sinus CT image data of patients with different genders, ages and disease course stages; dividing the paranasal sinus CT image data into a training verification set and a test set, and training an nnUNetV2 model based on the training verification set to obtain a chronic sinusitis region-of-interest segmentation model; and performing full-automatic segmentation of chronic sinusitis on CT image data of a target patient through the segmentation model of the region-of-interest of chronic sinusitis. The problems of low segmentation efficiency, high subjectivity and difficulty in precise segmentation of the chronic sinusitis region-of-interest caused by high complexity and individual variability of the anatomical structure of the nasal cavity and the paranasal sinus can be solved.

Description

Technical Field

[0001] This application relates to the field of smart healthcare, and more specifically, to a fully automated segmentation method and related equipment for chronic sinusitis based on deep learning. Background Technology

[0002] Chronic rhinosinusitis (CRS) is a common chronic inflammatory disease of the upper respiratory tract, with an epidemiological incidence rate as high as 8%. The highly complex and individual-variable anatomy of the nasal cavity and sinuses poses a significant challenge to the diagnosis and treatment of CRS. The nasal cavity and sinuses consist of four main cavities: the frontal sinuses, ethmoid sinuses, maxillary sinuses, and sphenoid sinuses, which are interconnected through the ostiomeatal complex. The interplay of anatomical variations and pathological changes makes accurately identifying and segmenting lesions in chronic rhinosinusitis from CT images a highly challenging clinical problem. These anatomical variations exhibit diverse morphologies on CT images and often overlap with inflammatory lesions, making accurate delineation of lesion boundaries a clinical challenge. Even experienced radiologists struggle to achieve precise segmentation of regions of interest in chronic rhinosinusitis when faced with CT images interwoven with complex anatomical variations and inflammatory lesions, posing a serious challenge to current medical image processing techniques. Summary of the Invention

[0003] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. The summary section of this invention is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0004] To address the problems of low efficiency, high subjectivity, and difficulty in accurate segmentation of regions of interest (ROIs) in chronic sinusitis caused by the high complexity and individual variability of the nasal cavity and sinuses, this invention proposes, in a first aspect, a fully automated segmentation method for chronic sinusitis based on deep learning. The method includes:

[0005] Acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages, and disease stages;

[0006] The sinus CT image data is divided into a training validation set and a test set, and the nnUNetV2 model is trained based on the training validation set to obtain a region of interest segmentation model for chronic sinusitis.

[0007] The region of interest segmentation model for chronic sinusitis is used to automatically segment CT image data of target patients for chronic sinusitis.

[0008] Optionally, a nonlocal means denoising algorithm is used to denoise the sinus CT image data of the patients with chronic sinusitis;

[0009] Normalize the grayscale values ​​of sinus CT image data to the [0,1] interval;

[0010] The images in the sinus CT image data are cropped according to the anatomical extent of the nasal cavity and sinuses, and irrelevant background areas are removed.

[0011] Perform at least one of the following data augmentation operations on the cropped image data: random rotation, scaling, translation, and brightness and contrast adjustment.

[0012] Optionally, the sinus CT image data sources include multiple clinical data centers. The step of dividing the sinus CT image data into training and validation sets and a test set, and training the nnUNetV2 model based on the training and validation sets to obtain a region of interest segmentation model for chronic sinusitis, includes:

[0013] The sinus CT image data from the main clinical data centers of multiple clinical data centers are randomly divided into an internal training validation set and an internal test set according to a preset ratio.

[0014] The nnUNetV2 model was pre-trained using a publicly available medical image dataset to enable the nnUNetV2 model to learn basic image features;

[0015] The model parameters of the nnUNetV2 model are fine-tuned based on the internal training and validation set.

[0016] Optional, also includes:

[0017] The trained nnUNetV2 model was tested on the internal test set and external test sets in other data centers.

[0018] Optionally, testing the trained nnUNetV2 model on the internal test set and external test sets in other data centers includes:

[0019] Input the CT image data from the test set into the trained nnUNetV2 model to obtain the segmentation results of the region of interest output by the model;

[0020] The segmentation results of the output region of interest are evaluated based on the Dice coefficient or IoU index to assess the segmentation performance and accuracy of the trained nnUNetV2 model.

[0021] Optional, also includes:

[0022] Statistical analysis was performed on the evaluation results from different data centers to assess the generalization ability of the trained nnUNetV2 model in different clinical settings.

[0023] Optionally, fine-tuning the model parameters of the nnUNetV2 model based on the internal training and validation set includes:

[0024] The nnUNetV2 model is trained on the internal training and validation set for a preset number of epochs to fine-tune the model parameters. The preset number of epochs is greater than or equal to 1000.

[0025] Secondly, this invention also proposes a fully automated segmentation device for chronic sinusitis based on deep learning, comprising:

[0026] The acquisition unit is used to acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages and disease stages.

[0027] The training unit is used to divide the sinus CT image data into a training validation set and a test set, so as to train the nnUNetV2 model based on the training validation set to obtain a region of interest segmentation model for chronic sinusitis.

[0028] The segmentation unit is used to perform fully automatic segmentation of the target patient's CT image data for chronic sinusitis using the region of interest segmentation model for chronic sinusitis.

[0029] Thirdly, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the deep learning-based fully automated segmentation method for chronic sinusitis as described in any of the first aspects above.

[0030] Fourthly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the deep learning-based fully automatic segmentation method for chronic sinusitis according to any one of the preceding claims of the first aspect.

[0031] In summary, the deep learning-based fully automated segmentation method for chronic sinusitis proposed in this application acquires sinus CT image data from patients with chronic sinusitis, including data from patients of different genders, ages, and disease stages. The sinus CT image data is divided into training and validation sets and a test set. An nnUNetV2 model is trained based on the training and validation sets to obtain a region of interest (ROI) segmentation model for chronic sinusitis. This RIO segmentation model is then used to automatically segment the CT image data of the target patient for chronic sinusitis. Thus, nnUNetV2, through automatic network adaptation mechanisms (such as channel number adjustment and receptive field control), can adapt to variations in the number of ethmoid sinus cells and morphological variations such as Onodi cells and Haller cells. Its 3D processing capabilities compensate for the lack of tomographic information in the Z-axis direction in 2D models. Although the original Regions of Interest (ROIs) used to train segmentation models are mostly obtained using traditional image processing methods based on grayscale thresholding and edge detection, these methods still suffer from incomplete segmentation or discontinuous boundaries in areas with complex morphology, significant anatomical variations, or slow transitions in lesion edges. Deep networks, however, can make distinctions by learning subtle differences in grayscale distribution from a large number of samples. For example, in the anterior sellar type of sphenoid sinus, the boundary between mucosal thickening and undeveloped cavities is blurred, but the model can identify abnormal signals by extracting contextual features and comparing changes in adjacent slices. The model automatically segments the regions of interest (ROIs) for each part of the sinuses, replacing manual delineation that takes more than 30 minutes. It provides automated ROI output, offering quantitative data for preoperative navigation, postoperative assessment, and quantification of inflammation severity, achieving high efficiency while maintaining accuracy and objectivity. Through multi-source training samples and data augmentation strategies, the model can adapt to image quality differences caused by different hospital CT scanning equipment and parameter settings, and can also generalize to cases of acute sinusitis or postoperative recurrence.

[0032] The present invention provides a fully automated segmentation method for chronic sinusitis based on deep learning. Other advantages, objectives and features of the present invention will be apparent in part from the following description, and in part from the understanding of those skilled in the art through study and practice of the invention. Attached Figure Description

[0033] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0034] Figure 1 A schematic diagram of a fully automated segmentation method for chronic sinusitis based on deep learning, provided for an embodiment of this application;

[0035] Figure 1a A comparative diagram illustrating the effect of region of interest segmentation in chronic sinusitis provided in the embodiments of this application;

[0036] Figure 1b The Dice coefficient diagram provided for embodiments of this application;

[0037] Figure 1c IoU index diagram provided for embodiments of this application

[0038] Figure 2 A schematic diagram of a fully automated segmentation device for chronic sinusitis based on deep learning, provided for an embodiment of this application;

[0039] Figure 3 This is a schematic diagram of a fully automated segmentation electronic device for chronic sinusitis based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0040] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0041] Understandably, the anatomical structure of the nasal cavity and sinuses is highly complex and varies greatly from person to person, posing a significant challenge to the diagnosis and treatment of CRS. The nasal cavity and sinuses consist of four main sinus cavities: the frontal sinus, ethmoid sinus, maxillary sinus, and sphenoid sinus, all interconnected via the ostiomeatal complex. The ethmoid sinus, being the most anatomically complex region, contains 3-18 air cells, with irregular shapes and significant differences in size. Some air cells are separated from the orbit and skull base only by a thin bony wall, as thin as 0.1-0.3 mm. On CT images, the boundaries of these air cells have similar grayscale values ​​to inflamed tissue, increasing the difficulty of lesion identification. Variations in the ethmoid sinus are not only reflected in the number and shape of air cells but also in their relationship with surrounding structures. For example, the incidence of Onodi air cells (the lateralmost air cells of the posterior ethmoid sinus) is approximately 10%-15%. Overdevelopment of these cells can lead to bone defects in the optic canal and internal carotid canal, increasing the risk of optic nerve injury and internal carotid artery rupture during endoscopic sinus surgery. The variations in the frontal sinuses are also significant, with substantial individual differences in their development. Approximately 20% of individuals have unilateral or bilateral undeveloped frontal sinuses, while well-developed frontal sinuses can have a volume of 10-30 ml. Furthermore, their shapes are diverse, ranging from triangular, rectangular, to irregular. The location and size of the frontal sinus openings also vary; in some individuals, the openings can be as narrow as 1-2 mm, easily leading to poor drainage and persistent inflammation. The maxillary sinus, as the largest sinus cavity, has an average volume of approximately 15-30 ml, but significant individual differences exist. Variations mainly manifest in sinus septa, sinus wall bone thickness, and the location of the sinus openings. About 5%-10% of individuals have bony septa within the maxillary sinus, dividing the sinus cavity into multiple small chambers, localizing inflammation and increasing the difficulty of treatment. Furthermore, the maxillary sinus floor is closely related to the tooth roots. In some patients, the maxillary sinus floor may be significantly concave due to excessive root growth, and in some cases, the tooth roots may even protrude into the sinus cavity. This anatomical variation easily leads to odontogenic sinusitis. The sphenoid sinus is located deep in the skull base, and its shape and size vary considerably. According to anatomical classification, the sphenoid sinus can be divided into thyroid type (approximately 2%), presellar type (approximately 12%), and saddle type (approximately 86%). Thyroid type sphenoid sinuses are underdeveloped, with a small cavity; presellar type sphenoid sinuses have insufficient pneumatization, and the cavity does not extend to the anterior sella turcica; saddle type sphenoid sinuses have good pneumatization and are closely related to the sella turcica. Variations in the sphenoid sinus significantly affect transnasal sphenoid sinus surgery. When sphenoid sinus pneumatization is poor or its adjacent relationship with the internal carotid artery or optic canal is abnormal, the risk of damage to important structures during surgery increases significantly. The ostiomeatal complex, as a key channel for nasal and sinus ventilation and drainage, has a cross-sectional area of ​​less than 1 cm². 2Mucosal edema, polyp hyperplasia, and other pathological changes can easily lead to obstruction in this area, forming a vicious cycle of inflammation. Clinical studies have shown that more than 80% of CRS patients have anatomical abnormalities or pathological changes in the sinus ostium-nasal meatus complex. These anatomical variations and pathological changes are intertwined, making accurate identification and segmentation of the lesion area of ​​chronic sinusitis from CT images a highly challenging clinical problem. In addition, nasal septum deviation further complicates the diagnosis and treatment of the disease, with an incidence rate as high as 70% in the population. It can lead to changes in nasal airflow dynamics, causing abnormal local mucosal blood perfusion and immune microenvironment imbalance, thereby inducing an inflammatory response. The incidence of vesicular middle turbinate (middle turbinate pneumatization) is about 20%-30%. Excessively pneumatized middle turbinate can occupy nasal space and obstruct sinus drainage. The presence of Haller's cells (superior wall air cells of the maxillary sinus) can lead to stenosis of the maxillary sinus ostium. Studies have shown that individuals with Haller's cells have a 40% higher incidence of CRS than those without this variation. These anatomical variations exhibit diverse morphologies on CT images and often overlap with inflammatory lesions, making accurate delineation of lesion boundaries a clinical challenge. Even experienced radiologists struggle to achieve precise segmentation of regions of interest in chronic sinusitis when faced with CT images interwoven with complex anatomical variations and inflammatory lesions, posing a significant challenge to current medical image processing technologies.

[0042] Currently, region-of-interest segmentation for chronic sinusitis mainly relies on semi-automatic segmentation methods. These methods typically require a specialist to manually delineate the lesion area, followed by segmentation using image processing software. This approach is not only time-consuming and labor-intensive, but also suffers from inconsistent and inaccurate segmentation results due to variations in individual physician experience and subjective judgment. This severely restricts the efficiency of clinical diagnosis of chronic sinusitis and the reliability of research data.

[0043] In the segmentation of other diseases, semi-automatic segmentation algorithms such as region growing, level set methods, and thresholding are widely used. These algorithms require users to provide certain initial information, such as seed points or contour curves, and then the algorithm expands the region or evolves the contour based on the image's grayscale, texture, and other features. However, for areas with complex anatomical structures and blurred lesion boundaries, such as the nasal cavity and sinuses, in practice, manual drawing is often used to determine the edges based on the thresholding algorithm. This semi-automatic segmentation algorithm also faces the problems of cumbersome operation, low efficiency, and reliance on manual intervention.

[0044] In recent years, deep learning technology has made groundbreaking progress in the field of fully automated medical image segmentation, with various innovative models emerging and demonstrating excellent performance. The classic U-Net model, as a representative of fully convolutional networks, effectively extracts semantic information and detailed features from images through its encoder-decoder structure and skip connection mechanism. It performs exceptionally well in 2D medical image segmentation tasks, especially in the segmentation of relatively well-defined structures such as skin lesions and retinal vessels, achieving a Dice coefficient of over 0.9. Its 3D extended version, 3D-U Net, further extends the network structure to 3D space, fully utilizing the volumetric data information of medical images. It has achieved significant results in the segmentation of 3D structures such as brain tumors and lung nodules. On the public dataset BraTS, 3D-U Net can maintain a stable Dice coefficient of around 0.85 for the overall segmentation of brain tumors.

[0045] nnU-Net is an adaptive medical image segmentation framework that automatically adjusts the network architecture, data augmentation strategies, and training parameters according to the characteristics of the dataset. It has achieved top rankings in multiple medical image segmentation challenges, with an average Dice coefficient of 0.92 in the liver segmentation task.

[0046] However, despite the excellent performance of the aforementioned algorithms in many medical image segmentation tasks, they still face significant challenges in the segmentation of chronic sinusitis. Nasal diseases possess unique anatomical complexity. The anatomical structure of the nasal cavity and sinuses includes multiple sinus cavities such as the frontal sinus, ethmoid sinus, maxillary sinus, and sphenoid sinus, which are interconnected through a narrow ostiomeatal complex, and there are significant anatomical variations between individuals. Taking the ethmoid sinus as an example, the number of air cells varies from 3 to 18, with irregular shapes, and some air cells are separated from the orbit and skull base only by a thin bony wall. On CT images, the grayscale difference between the boundaries of these air cells and inflamed tissue is minimal. At the same time, the pathological manifestations of chronic sinusitis are diverse. The grayscale values ​​of lesions such as mucosal thickening and polyp formation caused by inflammation highly overlap with those of normal tissue, making it difficult for conventional segmentation algorithms based on grayscale and texture features to accurately distinguish them.

[0047] Existing semi-automatic segmentation methods are difficult to apply in the clinical diagnosis and research of chronic sinusitis due to their time-consuming and labor-intensive nature, making clinical practice extremely challenging. An experienced radiologist typically spends half an hour processing sinus CT images of a single chronic sinusitis patient using traditional semi-automatic segmentation methods. From a research perspective, semi-automatic segmentation methods also severely restrict research progress. Large-scale clinical studies of chronic sinusitis often require processing hundreds or even thousands of CT image datasets. For example, in a multicenter study involving 500 patients, using semi-automatic segmentation methods would require 1000-2000 hours for image segmentation alone, not even considering the time spent on data annotation verification and correction. Such a massive time investment significantly delays research projects, making it difficult to complete data processing and analysis within the expected timeframe.

[0048] Meanwhile, because the segmentation process relies heavily on manual operation, differences in segmentation standards and operating habits exist between different research teams or even among different doctors within the same team, leading to insufficient consistency and comparability of the data. Furthermore, doctors' subjective judgments of the anterior and posterior boundaries of the sinuses vary. Due to the lack of unified and objective segmentation standards, different doctors exhibit significant differences in their understanding and judgment of chronic sinusitis lesion areas in the same CT image. This is directly reflected in the subjective differences even in the traditional CT Lund-Mackey scoring system among different doctors. These differences are not only reflected in the definition of lesion boundaries but also in the assessment of lesion severity. For example, for areas with mild mucosal thickening, some doctors may classify the entire area as a lesion, while others tend to mark only the more significantly thickened parts. Moreover, even the same doctor may experience fluctuations in segmentation results due to fatigue, distraction, and other factors when performing segmentation at different times and under different working conditions. These problems can directly affect the quality of the data and the credibility of the research conclusions, hindering in-depth research on chronic sinusitis.

[0049] To address the challenges of low efficiency, high subjectivity, and difficulty in precise segmentation of regions of interest (ROIs) in chronic sinusitis due to the highly complex and individual-variable anatomical structure of the nasal cavity and sinuses, please refer to [link to relevant documentation]. Figure 1 This is a schematic diagram of a fully automated segmentation method for chronic sinusitis based on deep learning provided in an embodiment of this application, which may specifically include steps S110 to S130.

[0050] S110, Acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages, and disease stages. The sinus CT image data may also include sinus CT image data of patients of different genders, ages, disease stages, medical institutions, and examination equipment.

[0051] S120, the sinus CT image data is divided into a training validation set and a test set, and the nnUNetV2 model is trained based on the training validation set and multiple cross-validations are performed to obtain a region of interest segmentation model for chronic sinusitis.

[0052] S130, The target patient's CT image data is automatically segmented for chronic sinusitis using the region of interest segmentation model for chronic sinusitis.

[0053] Understandably, nnUNetV2 is a deep learning model based on a 3D fully convolutional network, whose architecture consists of an encoder and a decoder. The encoder uses a series of convolutional and pooling layers to progressively extract high-level semantic features from the image, reducing image resolution. The decoder, through deconvolutional layers and skip connections, fuses the features extracted by the encoder with the low-level features of the original image to restore image resolution and achieve accurate pixel-level segmentation. Regarding model parameter settings, the basic number of channels is set to 32, and the number of channels gradually increases proportionally as the network depth increases to meet the needs of feature extraction at different levels. The convolutional kernel size is 3×3×3, ensuring feature extraction capability while reducing computational cost. The model uses the Adam optimizer for parameter updates and is trained using 5-fold cross-validation, with the learning rate dynamically adjusted during the process to improve convergence speed and training stability. The loss function is a weighted sum of Dice loss and cross-entropy loss. Dice loss measures the overlap between the segmentation result and the true label, while cross-entropy loss optimizes the model's classification performance. By adjusting the weights of both, the model's segmentation accuracy is effectively improved. To address the challenges of significant anatomical variations and blurred lesion boundaries in CT images of patients with chronic sinusitis, a multi-structure strategy under the nnUNetV2 framework was adopted in the training process, training four typical filtering architectures: 2D, 3D_lowres, 3D_fullres, and 3D_cascade_fullres. The 2D network can quickly capture detailed textures in axial slices, suitable for superficial lesion recognition; the 3D_lowres network enhances contextual modeling capabilities through downsampling, adapting to anatomical inference of large-volume structures; the 3D_fullres network preserves high-resolution voxel information, strengthening the learning of local edge structures; and 3D_cascade_fullres, based on a coarse-to-fine strategy, first coarsely segments and then refines the prediction results, improving the ability to express complex multi-celled ethmoid sinus structures. Training results from different architectures were integrated and analyzed. By comparing the segmentation accuracy, stability, and inference efficiency of different architectures, the prediction results of 2D and 3D_fullres were selected for ensemble-2d-3d_fullres integration to improve multi-scale structural representation and detail restoration capabilities. The integration process employed a voxel-level averaging fusion strategy, weighting and merging the outputs of different networks to obtain a final chronic sinusitis segmentation model that combines boundary fineness with global stability.Specifically addressing the issue of inconsistent and incomparable data due to the highly manual segmentation process, which leads to differences in segmentation standards and operating habits among different research teams or even different doctors within the same team, and the subjective judgments of doctors regarding the anterior and posterior boundaries of the sinuses, this method enhances the consistency and reproducibility of clinical assessments. It significantly reduces labeling bias caused by differences in doctor experience, operating habits, or subjective perception, particularly resolving the common problem of subjective ambiguity in the manual segmentation of structures such as the anterior and posterior boundaries of the sinuses, the posterior chamber of the ethmoid sinus, and the floor of the maxillary sinus. By learning boundary patterns on a large amount of multi-center data, the model unifies structural definitions and spatial attributions, thereby improving data consistency and comparability.

[0054] For example, a large amount of representative sinus CT image data from patients with chronic sinusitis needs to be acquired. This data should cover different genders, age groups (children, adults, and the elderly), different disease stages (early stage, progressive stage, postoperative recurrence, etc.), and different anatomical variations (such as Haller's cells, Onodi's cells, bullous middle turbinate, nasal septum deviation, etc.) to ensure sample diversity and generalization. The acquired CT image data needs to undergo standardized preprocessing operations, such as spatial resampling of the images (unified to isovoxel resolution, such as 1×1×1mm). 3 The process included intensity normalization (mapping CT grayscale values ​​to the [0,1] or [-1,1] range to remove interference from grayscale shifts between different devices), and background removal (removing extracranial tissues through coarse segmentation or threshold filtering). CT scan parameters were uniformly set to slice thickness 1-2mm, matrix 512×512, window width 2000-2500HU, and window level 300-500HU to ensure image quality met subsequent processing and analysis requirements. Based on this, voxel-level manual annotation was performed, and a gold standard mask for sinus cavities and inflammatory areas was drawn by a senior ENT specialist using 3D visualization tools (such as ITK-SNAP and 3D Slicer). This included normal sinus cavities and lesion areas (mucosal thickening, polyps, effusion, etc.) in various regions such as the frontal sinus, ethmoid sinus, sphenoid sinus, maxillary sinus, and nasal cavity. Subsequently, all data were divided into training-validation and test sets according to a preset ratio to ensure that different variant samples were evenly distributed across each subset. For example, patient data from the main data center can be randomly divided into a training validation set and an internal test set in an 8:2 ratio.

[0055] For example, training can be based on the nnUNetV2 model. nnUNetV2 is an adaptive medical image segmentation framework that can automatically adjust the network structure (such as depth, number of channels, and kernel size), data augmentation strategies (such as 3D affine transformation, elastic deformation, and random noise injection), and optimizer and training hyperparameter settings according to the spatial size, label distribution, and anatomical complexity of the input data. During the training phase, the training set data can be input into the network in patch form for forward propagation. The network adopts a 3D encoder-decoder structure and combines a skip connection mechanism to extract global semantic information and local detail features. At the same time, a deep supervision mechanism is introduced to enhance the segmentation effect on multi-scale targets. The Dice loss and cross-entropy loss functions are used to jointly optimize the model parameters, maximizing the solution to the difficult problems of gray-level overlap and blurred boundaries between sinus lesions and normal tissues. During training, the model learns the differences in the number, morphology, and degree of pneumatization of ethmoid sinus air cells among patients through multiple epoch iterations, as well as the pathological manifestations corresponding to features such as frontal sinus hypoplasia, maxillary sinus bony septum, and odontogenic variations. It is dynamically evaluated on the validation set and overfitting is prevented through an early stopping strategy. Finally, the best-performing weighted model is saved.

[0056] For example, CT images of the target patient can be input into a trained nnUNetV2 model for fully automated segmentation. The specific operation involves preprocessing the patient's 3D CT images according to the model's input specifications, then inputting them into the model for inference. The model outputs voxel-level multi-class prediction maps, such as 0 for normal airways, 1 for mucosal thickening, and 2 for polyps, automatically forming a complete segmentation mask for chronic sinusitis lesions. To improve clinical usability, the system can also provide 3D visualization of the segmentation results. Through transparent rendering and interactive slice browsing, doctors can intuitively identify the extent, location, and relationship to anatomical structures of the lesions. When combined with a preoperative planning and navigation system, the segmentation results can be further projected into the intraoperative image stream, assisting doctors in safely and accurately completing endoscopic sinus surgery.

[0057] Understandably, the aforementioned method, based on the adaptive segmentation capabilities of deep neural networks and their high robustness to complex sinus anatomical variations, achieves rapid, stable, and accurate extraction of lesion regions. Compared to traditional manual segmentation, it improves time efficiency by tens of times and maintains high-precision identification even in high-risk areas such as Onodi air cells, sphenoid sinus saddle shapes, and Haller air cells. This improves lesion identification accuracy and reduces the rate of missed diagnoses, especially when polypoid lesions are similar in grayscale to normal bone walls. It also enhances the consistency and repeatability of clinical assessment, avoiding errors introduced by the subjectivity of manual annotation. Furthermore, it significantly reduces physician workload, providing a technical foundation for large-scale imaging studies of chronic sinusitis, preoperative assessment, and model-assisted diagnosis and treatment. In addition, by incorporating the learning of rare variant structures during model training, it improves adaptability to individual anatomical differences, possesses cross-population and cross-device generalization capabilities, and has extremely high clinical application value.

[0058] In summary, the deep learning-based fully automated segmentation method for chronic sinusitis provided in this application acquires sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages, and disease stages; divides the sinus CT image data into training and validation sets and test sets, and trains the nnUNetV2 model based on the training and validation sets to obtain a region of interest (ROI) segmentation model for chronic sinusitis; and uses the ROI segmentation model to automatically segment the CT image data of the target patient for chronic sinusitis. Thus, nnUNetV2, through automatic network adaptation mechanisms (such as channel number adjustment and receptive field control), can adapt to differences in the number of ethmoid sinus cells and morphological variations such as Onodi cells and Haller cells. Its 3D processing capability compensates for the lack of tomographic information in the Z-axis direction of the 2D model. Although the original Regions of Interest (ROIs) used to train segmentation models are mostly obtained using traditional image processing methods based on grayscale thresholding and edge detection, these methods still suffer from incomplete segmentation or discontinuous boundaries in areas with complex morphology, significant anatomical variations, or slow transitions in lesion edges. Deep networks, however, can make distinctions by learning subtle differences in grayscale distribution from a large number of samples. For example, in the anterior sellar type of sphenoid sinus, the boundary between mucosal thickening and undeveloped cavities is blurred, but the model can identify abnormal signals by extracting contextual features and comparing changes in adjacent slices. It automatically segments the regions of interest in various parts of the sinuses, replacing manual delineation that takes more than 30 minutes; it provides automated ROI output, providing quantitative evidence for preoperative navigation, postoperative assessment, and quantification of inflammation severity, achieving high efficiency while being as accurate and objective as possible. Through multi-source training samples and data augmentation strategies, the model can adapt to image quality differences caused by different hospital CT scanning equipment and parameter settings, and can also generalize to cases of acute sinusitis or postoperative recurrence.

[0059] It should be noted that the above-mentioned approach targets the global challenge of reducing the recurrence rate of CRS. Addressing key bottlenecks such as the unclear immunopathological characteristics, ambiguous pathogenesis, and lack of precise treatment strategies for CRS, it has established a precision diagnosis and treatment system for CRS through persistent and systematic research. Innovative technologies such as the "diagnostic criteria for eosinophilic CRS," "non-invasive subtyping diagnosis and efficacy prediction technology for CRS," and "precision diagnosis and treatment strategies for CRS" have been widely promoted and applied. These systems have been widely used in hospitals, serving over 1 million patients cumulatively, reducing the one-year recurrence rate of CRS from over 30% to below 16%. The achievements have led to a leapfrog development of CRS from a "uniform diagnosis and treatment" model to a "precision subtyping diagnosis and treatment" model, providing an important paradigm for the diagnosis and treatment of chronic respiratory diseases globally.

[0060] In some examples, it also includes:

[0061] The sinus CT image data of the patients with chronic sinusitis were denoised using a nonlocal mean denoising algorithm.

[0062] Normalize the grayscale values ​​of sinus CT image data to the [0,1] interval;

[0063] The images in the sinus CT image data are cropped according to the anatomical extent of the nasal cavity and sinuses, and irrelevant background areas are removed.

[0064] Perform at least one of the following data augmentation operations on the cropped image data: random rotation, scaling, translation, and brightness and contrast adjustment.

[0065] Understandably, in order to improve the effectiveness of model training and segmentation accuracy, multiple preprocessing and enhancement operations can be introduced before the data is formally input into the deep learning network, thereby enhancing the model's noise robustness, anatomical adaptability, and training stability on CT images of chronic sinusitis.

[0066] For example, a nonlocal mean denoising algorithm can be used to process the original sinus CT image data. This algorithm finds a pixel region in the entire image that is similar to the pixel in the image and performs a weighted average, which effectively preserves the anatomical edge details and removes interference such as random Gaussian noise and X-ray artifacts in the image. It is especially suitable for the clear presentation of the boundary area between thin bone layers and low grayscale tissue in CT images. This has a significant effect on segmenting blurred structures such as the sphenoid sinus wall and ethmoid sinus air cells. After denoising, grayscale deviations caused by different scanning instruments and scanning protocols are eliminated.

[0067] For example, image intensity can be normalized, linearly mapping all voxel gray values ​​to the [0,1] interval. This not only accelerates model convergence but also reduces training instability caused by differences in input distribution. Subsequently, to focus on the nasal cavity and sinus regions, an anatomical location-based cropping strategy is used to spatially crop the 3D CT image, removing irrelevant areas such as the skull and maxillofacial soft tissues, while retaining the ROI from the nasal vestibule to the posterior wall of the sphenoid sinus. This cropping not only reduces background redundancy but also improves the spatial resolution of the model's focus, thereby enhancing the model's perceptual ability and computational efficiency. Based on the above processing, To enhance the model's generalization ability to anatomical variations, acquisition angles, and lighting conditions, at least one data augmentation strategy is introduced into the cropped image data. This includes random rotation (simulating different head tilt positions), scaling (adapting to patients of different body types or developmental stages), translation (enhancing the model's spatial tolerance), brightness adjustment (simulating exposure differences), and contrast enhancement (strengthening structural texture boundaries). These enhancement operations are dynamically applied with a certain probability in each training iteration, ensuring that the model encounters more data distributions with real-world variability during the learning process, thereby improving the model's ability to recognize and segment complex real-world scenes. The combined preprocessing steps significantly improve the model's ability to distinguish lesion regions and anatomical boundaries, enhance training stability and final segmentation accuracy, and demonstrate superior performance, particularly in chronic sinusitis images with blurred anatomical structures and significant gray-level overlap of lesions.

[0068] In some examples, the sinus CT image data sources include multiple clinical data centers. The step of dividing the sinus CT image data into training and validation sets and a test set, and then training the nnUNetV2 model based on the training and validation sets to obtain a region of interest segmentation model for chronic sinusitis, includes:

[0069] The sinus CT image data from the main clinical data centers of multiple clinical data centers are randomly divided into an internal training validation set and an internal test set according to a preset ratio.

[0070] The nnUNetV2 model was pre-trained using a publicly available medical image dataset to enable the nnUNetV2 model to learn basic image features;

[0071] The model parameters of the nnUNetV2 model are fine-tuned based on the internal training and validation set.

[0072] Understandably, to improve the model's generalization ability, robustness, and clinical adaptability, the sinus CT image data is not limited to a single institution but is collected from multiple clinical data centers, covering different equipment models, scanning protocols, pathological features, and patient populations. To fully utilize these heterogeneous data sources while ensuring the model's high performance across various clinical scenarios, a phased training strategy is employed to construct the final region of interest segmentation model for chronic sinusitis.

[0073] For example, data from the primary clinical data center with the most stable data quality and annotation quality among multiple clinical data centers can be used as the basic training resource. This data can then be randomly divided according to a preset ratio (e.g., 80% for training validation and 20% for internal testing) to generate internal training validation and test sets. During the partitioning process, the balanced distribution of various anatomical variations, gender, age group, CT slice thickness, and other factors should be fully considered to avoid model bias towards a particular data structure or individual characteristics.

[0074] For example, the nnUNetV2 model can be pre-trained using publicly available medical image datasets (such as Head-Neck CTSegmentation, MICCAI structured dataset, etc.). During the pre-training stage, the model learns basic image construction rules, edge structure recognition capabilities, and anatomical region perception capabilities through large-scale general CT images, thereby acquiring general feature extraction capabilities. This stage mainly uses unlabeled or weakly labeled data for self-supervised or supervised learning, resulting in a model with better initial parameter configurations, reducing the difficulty of subsequent training convergence.

[0075] For example, after completing the basic pre-training, the model can be fine-tuned based on the internal training and validation set constructed above. This process is the supervised learning phase, which uses labeled CT images of chronic sinusitis for fully convolutional training to specifically optimize the model's ability to recognize lesion regions (such as mucosal thickening, polypoid hyperplasia, sinus effusion, etc.) at the voxel level. The fine-tuning phase typically uses a low learning rate and enables transfer learning mechanisms to refine the output's ability to identify the segmentation boundaries of target anatomical structures while maintaining the stability of the original model's feature extraction structure.

[0076] Understandably, the above method addresses the issue of differences in image quality and structural representation across different data centers. It also fully leverages the feature generalization capabilities of publicly available large-scale data and the task-specific advantages of locally sourced high-quality labeled data, thereby significantly improving the model's accuracy and robustness in image segmentation tasks for chronic sinusitis. Particularly when dealing with highly variable regions such as Onodi air cells, lesions near the skull base, and bony septa of the maxillary sinus, the model demonstrates stronger boundary-capturing ability and cross-scenario adaptability, further enhancing its reliability and universality in clinical applications.

[0077] In some examples, it also includes:

[0078] The trained nnUNetV2 model was tested on the internal test set and external test sets in other data centers.

[0079] Understandably, in order to comprehensively evaluate the generalization performance and practical application value of the trained nnUNetV2 model in different clinical scenarios, the method also includes independently testing the trained segmentation model on internal and external test sets.

[0080] For example, the internal test set refers to a subset of data from the main clinical data center that was not involved in the model training process. Its structure and distribution are relatively similar to the training set. It is mainly used to evaluate the model's ability to fit similar data and its segmentation accuracy. On the other hand, the external test set comes from other clinical data centers that were not involved in the model training. These data often have more obvious anatomical variability, differences in acquisition protocols, image quality fluctuations, and even include image data acquired from cross-brand devices. The purpose is to test the model's adaptability and robustness in real heterogeneous clinical scenarios.

[0081] For example, during testing, these unseen data points are input into the trained nnUNetV2 model as 3D CT images. The output segmentation mask is obtained and compared with manually annotated gold standard labels. Evaluation metrics, including Dice coefficient, IoU (Intersection over Union), Hausdorff distance, and voxel accuracy, are calculated. On the internal test set, high Dice values ​​and low boundary errors reflect the model's accurate learning ability of lesion morphology; while performance on the external test set better verifies whether the model truly captures universal lesion features, rather than overfitting to the structural features of a specific data center.

[0082] For example, if the internal test set mainly comes from CT equipment of brand A used in the same hospital, the external test set may come from another hospital using equipment of brand B, and the incidence of Onodi air cells, Haller air cells, and vesicular middle turbinates may differ among patient groups. If the model can still maintain stable segmentation performance on such external test sets, it indicates that it has good adaptability to anatomical variations and cross-device transferability, truly meeting the general clinical application requirements.

[0083] Understandably, this testing process is not only a verification step for the training results, but also constitutes a key technical guarantee before the model is deployed. It helps to establish reasonable usage boundaries and confidence thresholds in the model-assisted diagnostic system, ensuring its safety and controllability in multi-center clinical applications.

[0084] In some examples, the step of testing the trained nnUNetV2 model on the internal test set and external test sets in other data centers includes:

[0085] Input the CT image data from the test set into the trained nnUNetV2 model to obtain the segmentation results of the region of interest output by the model;

[0086] The segmentation results of the output region of interest are evaluated based on the Dice coefficient or IoU index to assess the segmentation performance and accuracy of the trained nnUNetV2 model.

[0087] For example, to systematically and quantitatively evaluate the performance of the trained nnUNetV2 model, three-dimensional CT image data from the test set can be input one by one into the trained nnUNetV2 model. Based on its encoder-decoder structure and skip connection mechanism, the model automatically extracts features and reconstructs spatially the sinus structures and lesion areas in the input images, and outputs a voxel-level segmentation mask of the same dimension as the original image. The segmentation result indicates the region of interest (ROI), including the voxel location and boundary of lesion areas such as mucosal thickening, sinus effusion, polyp-like tissue, etc.

[0088] For example, based on the gold-standard manually annotated mask and the automatic segmentation results output by the model, DSC (Dice Similarity Coefficient) or IoU (Intersection over Union) is used as an evaluation metric to quantify the model's segmentation performance. The Dice coefficient measures the degree of spatial overlap between two regions, and its formula is:

[0089] Dice=2×|A∩B| / (|A|+|B|),

[0090] In this model, A represents the manually labeled region, and B represents the model's predicted region. A value closer to 1 indicates a more accurate segmentation result. The IoU index measures the ratio of the intersection to the union of the predicted and actual regions and is often used to evaluate multi-class segmentation scenarios. For example, if the model achieves a Dice value of over 0.88 for ethmoid sinus lesions on the internal test set, it indicates high accuracy in structurally complex and poorly defined regions. Further evaluation on the external test set of lesions such as Haller's air cells causing maxillary sinus ostium stenosis, if the Dice value remains around 0.80, it shows that the model not only works accurately on existing data but also maintains good robustness and generalization ability on unseen data.

[0091] Understandably, this evaluation process comprehensively verifies the segmentation capabilities of the nnUNetV2 model under different anatomical variations and pathological types of chronic sinusitis, effectively reflecting its usability and reliability as an auxiliary diagnostic tool in real clinical settings. Furthermore, the test results can serve as an important basis for subsequent model optimization, deployment evaluation, and risk control strategies, ensuring that the segmentation algorithm has rigorous technical support and data backing for widespread adoption.

[0092] For example, the nnUNetV2 model is trained on a training and validation set. The training process is performed on a high-performance computing platform, using distributed training techniques to accelerate the process. In each training round, the training data is divided into five batches, which are sequentially input into the model for training and cross-validation. During training, the model's training loss and validation metrics (such as the Dice index and IoU index) are monitored in real time, and model parameters and training strategies are adjusted based on the monitoring results. When the model achieves optimal performance on the internal validation set, the model parameters are saved as the final trained model.

[0093] For example, the external test set data came from data centers in 11 different regions across the country, covering CT scanning equipment from 8 mainstream brands including GE, Siemens, and Philips. The scanning parameters varied significantly (slice thickness ranged from 0.625 to 3.0 mm, tube voltage from 100 to 140 kV, and tube current from 80 to 300 mAs), and some data contained interference factors such as respiratory artifacts and metal artifacts. However, our system maintained stable segmentation performance in all cases. This strong generalization ability ensures the reliable application of this system in different medical environments, providing solid technical support for multi-center clinical research and telemedicine diagnosis.

[0094] In some examples, it also includes:

[0095] Statistical analysis was performed on the evaluation results from different data centers to assess the generalization ability of the trained nnUNetV2 model in different clinical settings.

[0096] Understandably, to verify the adaptability and stability of the trained nnUNetV2 model in a multi-source clinical environment, the method also includes: statistical analysis of the evaluation results from different data centers to systematically evaluate the model's cross-center generalization ability. This process is performed after the segmentation performance evaluation of the internal and external test sets is completed.

[0097] For example, the score distribution of the model output results and the gold standard label under multiple key evaluation indicators can be statistically analyzed on each data center, including the Dice coefficient, IoU index, and Hausdorff distance. The mean, standard deviation, maximum, and minimum values ​​of these indicators are calculated separately for each data center, and comparison plots or box plots are generated to visually reflect the consistency and fluctuation of the model's segmentation performance in different clinical scenarios. Further statistical methods such as ANOVA or paired t-tests can be used to determine whether the differences in indicators between different centers are statistically significant. If there are no significant differences, it indicates that the model has strong cross-center generalization ability; if there are significant differences, it suggests the need to further introduce domain adaptation strategies, enhance the training weights of specific center samples, or perform fine-tuning of the model across different centers. In addition, this statistical analysis can also identify which anatomical variations (such as Onodi air cells, Haller air cells, etc.) or pathological manifestations (such as mild mucosal thickening or polypoid nodules) are most likely to cause model segmentation errors, providing targeted optimization directions for subsequent model refinement and training set supplementation.

[0098] According to some embodiments, the fine-tuning of the model parameters of the nnUNetV2 model based on the internal training and validation set includes:

[0099] The nnUNetV2 model is trained on the internal training and validation set for a preset number of epochs to fine-tune the model parameters. The preset number of epochs is greater than or equal to 1000.

[0100] For example, after pre-training on a publicly available medical imaging dataset, the model can be transferred to a chronic sinusitis-related task environment and trained for at least 1000 epochs based on an internal training and validation set to fully adapt to the anatomical features and pathological manifestations in the target task. This internal training and validation set is constructed from annotated CT images from major clinical data centers, covering various types of chronic sinusitis lesions (such as mucosal thickening, polyps, and sinus effusion) and different anatomical variations (such as bullous middle turbinate, Onodi air cells, and nasal septum deviation), and its data distribution can represent the structural diversity and complexity in real clinical settings.

[0101] For example, the fine-tuning phase employs a transfer learning strategy. The initial model parameters are derived from the pre-training phase, preserving the basic image structure recognition capabilities. Fine-tuning then enhances the model's ability to extract sinus anatomy-specific features by progressively updating the higher-level parameters. In each epoch, training samples are normalized, cropped, and subjected to random data augmentation (including rotation, scaling, and changes in brightness and contrast) before being input into the nnUNetV2 model. The model extracts spatial contextual features of the image through a 3D convolutional network structure and then recovers the segmentation mask layer by layer through the decoder. A weighted combination of the Dice loss function and cross-entropy loss is introduced during training to maintain classification stability while optimizing the overlap of segmentation boundaries.

[0102] Understandably, training through 1000 or more iterations not only ensures the model's convergence but also improves its ability to identify difficult-to-segment regions (such as the ethmoid sinus-orbital junction and lesions on the anterior wall of the sphenoid sinus). Furthermore, to prevent overfitting, the model's performance is periodically evaluated using a validation set during training, employing strategies such as early stopping and adaptive learning rate decay to ensure training stability and generalization ability. The final output model exhibits high region-of-interest segmentation accuracy, effectively identifying different types of chronic sinusitis lesions in CT images, and demonstrates strong adaptability to anatomical variations, significantly improving the model's usability and safety in real-world clinical settings.

[0103] For example, such as Figure 1a As shown, by comparing the results with those of manual semi-automatic segmentation (left side), the fully automatic segmentation results of this invention (right side) are almost indistinguishable from manual segmentation in terms of visual morphology, with only slight differences at the edges. For example, the boundaries of the ROI in the fully automatic segmentation result exhibit jagged undulations, while the boundaries of the ROI in the manual semi-automatic segmentation result are relatively smoother. Analysis of a large number of samples shows that the fully automatic segmentation results can accurately cover the lesion area and have extremely high visual consistency with the manual segmentation results, providing a reliable diagnostic reference for clinical applications.

[0104] For example, such as Figure 1b As shown, the Dice index plot graphically displays the distribution of the Dice index on the training-validation set, internal test set, and external test set. Figure 1c As shown, the IoU index plot displays the distribution of IoU indices on the training-validation set, internal test set, and external test set in a graphical format. Statistical analysis software (such as Python's statistical analysis libraries) can be used to analyze the evaluation results, calculate the mean, standard deviation, and other statistics of each evaluation indicator on different datasets, and perform inter-group comparisons and significance tests.

[0105] Please see Figure 2One embodiment of the fully automated segmentation device for chronic sinusitis based on deep learning in this application may include:

[0106] Acquisition unit 21 is used to acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages and disease stages.

[0107] Training unit 22 is used to divide the sinus CT image data into a training validation set and a test set, so as to train the nnUNetV2 model based on the training validation set to obtain a region of interest segmentation model for chronic sinusitis.

[0108] The segmentation unit 23 is used to perform fully automatic segmentation of the target patient's CT image data for chronic sinusitis using the region of interest segmentation model for chronic sinusitis.

[0109] In summary, the deep learning-based fully automated segmentation device for chronic sinusitis provided in this application acquires sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages, and disease stages; divides the sinus CT image data into training and validation sets and test sets, and trains the nnUNetV2 model based on the training and validation sets to obtain a region of interest (ROI) segmentation model for chronic sinusitis; and uses the ROI segmentation model to automatically segment the CT image data of the target patient for chronic sinusitis. Thus, nnUNetV2, through automatic network adaptation mechanisms (such as channel number adjustment and receptive field control), can adapt to differences in the number of ethmoid sinus cells and morphological variations such as Onodi cells and Haller cells. Its 3D processing capability compensates for the lack of tomographic information in the Z-axis direction of the 2D model. Although the original Regions of Interest (ROIs) used to train segmentation models are mostly obtained using traditional image processing methods based on grayscale thresholding and edge detection, these methods still suffer from incomplete segmentation or discontinuous boundaries in areas with complex morphology, significant anatomical variations, or slow transitions in lesion edges. Deep networks, however, can make distinctions by learning subtle differences in grayscale distribution from a large number of samples. For example, in the anterior sellar type of sphenoid sinus, the boundary between mucosal thickening and undeveloped cavities is blurred, but the model can identify abnormal signals by extracting contextual features and comparing changes in adjacent slices. It automatically segments the regions of interest in various parts of the sinuses, replacing manual delineation that takes more than 30 minutes; it provides automated ROI output, providing quantitative evidence for preoperative navigation, postoperative assessment, and quantification of inflammation severity, achieving high efficiency while being as accurate and objective as possible. Through multi-source training samples and data augmentation strategies, the model can adapt to image quality differences caused by different hospital CT scanning equipment and parameter settings, and can also generalize to cases of acute sinusitis or postoperative recurrence.

[0110] like Figure 3As shown, this application embodiment also provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements the steps of any of the above-described methods for fully automatic segmentation of chronic sinusitis based on deep learning.

[0111] Since the electronic device described in this embodiment is the device used to implement the fully automatic segmentation device for chronic sinusitis based on deep learning in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in this application embodiment. Therefore, how the electronic device implements the method in this application embodiment will not be described in detail here. Any device used by those skilled in the art to implement the method in this application embodiment is within the scope of protection of this application.

[0112] In practical implementation, when the computer program 311 is executed by the processor, it can achieve the following: Figure 1 Any of the corresponding implementation methods in the embodiments.

[0113] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] This application also provides a computer program product, which includes computer software instructions that, when executed on a processing device, cause the processing device to perform actions such as... Figure 1 The corresponding embodiment shows the process of fully automated segmentation of chronic sinusitis based on deep learning.

[0119] A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0122] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A fully automated segmentation method for chronic sinusitis based on deep learning, characterized in that, include: Acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages, and disease stages; The sinus CT image data is divided into a training validation set and a test set, and the nnUNetV2 model is trained based on the training validation set to obtain a region of interest segmentation model for chronic sinusitis. The region of interest segmentation model for chronic sinusitis is used to automatically segment CT image data of target patients for chronic sinusitis.

2. The method as described in claim 1, characterized in that, Also includes: The sinus CT image data of the patients with chronic sinusitis were denoised using a nonlocal mean denoising algorithm. Normalize the grayscale values ​​of sinus CT image data to the [0,1] interval; The images in the sinus CT image data are cropped according to the anatomical extent of the nasal cavity and sinuses, and irrelevant background areas are removed. Perform at least one of the following data augmentation operations on the cropped image data: random rotation, scaling, translation, and brightness and contrast adjustment.

3. The method as described in claim 1, characterized in that, The sinus CT image data comes from multiple clinical data centers. The process of dividing the sinus CT image data into training and validation sets and a test set, and then training the nnUNetV2 model based on the training and validation sets to obtain a region of interest segmentation model for chronic sinusitis, includes: The sinus CT image data from the main clinical data centers of multiple clinical data centers are randomly divided into an internal training validation set and an internal test set according to a preset ratio. The nnUNetV2 model was pre-trained using a publicly available medical image dataset to enable the nnUNetV2 model to learn basic image features; The nnUNetV2 model is trained on the internal training and validation set for a preset number of epochs to fine-tune the model parameters. The preset number of epochs is greater than or equal to 1000.

4. The method as described in claim 3, characterized in that, Also includes: The trained nnUNetV2 model was tested on the internal test set and external test sets in other data centers.

5. The method as described in claim 4, characterized in that, The step of testing the trained nnUNetV2 model on the internal test set and external test sets in other data centers includes: Input the CT image data from the test set into the trained nnUNetV2 model to obtain the segmentation results of the region of interest output by the model; The segmentation results of the output region of interest are evaluated based on the Dice coefficient or IoU index to assess the segmentation performance and accuracy of the trained nnUNetV2 model.

6. The method as described in claim 5, characterized in that, Also includes: Statistical analysis was performed on the evaluation results from different data centers to assess the generalization ability of the trained nnUNetV2 model in different clinical settings.

7. The method according to any one of claims 3 to 6, characterized in that, The fine-tuning of the model parameters of the nnUNetV2 model based on the internal training and validation set includes: The nnUNetV2 model is trained on the internal training and validation set for a preset number of epochs to fine-tune the model parameters. The preset number of epochs is greater than or equal to 1000.

8. A fully automated segmentation device for chronic sinusitis based on deep learning, characterized in that, include: The acquisition unit is used to acquire sinus CT image data of patients with chronic sinusitis, including sinus CT image data of patients of different genders, ages and disease stages. The training unit is used to divide the sinus CT image data into a training validation set and a test set, so as to train the nnUNetV2 model based on the training validation set to obtain a region of interest segmentation model for chronic sinusitis. The segmentation unit is used to perform fully automatic segmentation of the target patient's CT image data for chronic sinusitis using the region of interest segmentation model for chronic sinusitis.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program stored in the memory, implements the steps of the fully automated segmentation method for chronic sinusitis based on deep learning as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the fully automated segmentation method for chronic sinusitis based on deep learning as described in any one of claims 1-7.

Images