Nasal polyp automatic grading model training method, scoring method, system and equipment
By constructing an endoscopic image set to train the target detection model, and utilizing data augmentation and hyperparameter tuning, the problems of automation and accuracy in nasal polyp grading were solved, achieving automated nasal polyp grading and improving the robustness and accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-13
AI Technical Summary
Existing nasal polyp grading methods rely heavily on physicians' subjective experience, leading to inconsistent grading results. Furthermore, artificial intelligence technology cannot automatically grade the position and size of nasal polyps relative to key anatomical structures such as the middle and inferior turbinates.
By constructing an endoscopic image set containing nasal polyp rating information of different grades, a pre-set target detection model is trained. Data augmentation, cross-validation, and hyperparameter tuning are used to ensure that the model can stably learn the international standard for nasal polyp rating and output an automatic and accurate nasal polyp rating.
It achieves automated and accurate nasal polyp grading, overcomes the functional limitations of artificial intelligence technology in nasal polyp grading, improves the model's robustness to differences in nasal polyp morphology and location, and reduces subjective bias.
Smart Images
Figure CN121661413A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image analysis technology, and in particular to an automatic grading model training method, scoring method, system and device for nasal polyps. Background Technology
[0002] Chronic sinusitis is a prevalent disease worldwide, significantly impacting patients' quality of life. In clinical practice, otolaryngologists use nasal endoscopy to quantitatively assess polyp size based on endoscopic polyp severity scoring, evaluating the effectiveness of surgical and drug treatments for chronic sinusitis. Currently, the nasal polyp severity scoring uses internationally accepted standards, with a unilateral nasal cavity score ranging from 0 to 4 points.
[0003] Existing methods for grading nasal polyps have significant limitations. These methods heavily rely on physicians' subjective experience and visual judgment, often leading to inconsistencies in scores between different doctors. To ensure the reliability of research data, even simultaneous evaluation by two doctors is sometimes necessary, but this dual-interpretation still struggles to completely eliminate subjective bias and significantly increases workload. While some artificial intelligence technologies attempt to identify polyps within the nasal cavity, they are used to differentiate between normal mucosa, polyps, and tumors, making it difficult to achieve automatic grading of nasal polyps in patients with chronic sinusitis.
[0004] Therefore, there is a lack of existing technologies that can automatically and accurately classify nasal polyps. Summary of the Invention
[0005] To address the problem that existing artificial intelligence technologies cannot automatically perform NPS grading based on the position and size of polyps relative to key anatomical structures such as the middle and inferior turbinates, this invention proposes an automatic grading model training method for nasal polyps, including: Acquire endoscopic medical imaging data of the nose from multiple patients with chronic sinusitis; The nasal endoscopic medical image data is preprocessed to obtain an endoscopic image set containing scoring and labeling information for different grades of nasal polyps; The endoscopic image set is divided into a training set, a validation set, and a test set; The training set is subjected to data augmentation processing, and the pre-set target detection model is cross-validated and trained using the data augmented training set to obtain multiple target detection models to be used. The performance of the target detection model to be used is evaluated based on the test set, and a target detection model to be used is selected based on the performance evaluation results. Based on the validation set and the data-enhanced training set, the selected target detection model is trained and its hyperparameters are tuned to obtain an automatic nasal polyp grading model.
[0006] Furthermore, the nasal endoscopic medical image data is video image, and the preprocessing of the nasal endoscopic medical image data includes: A single frame image for nasal polyp grading is extracted from the video image, the single frame image containing the anatomical structures required for grading; When the polyp grade score is 0, 1 or 2, the single frame image includes the ostiomeatal complex structure, the middle meatus and the lower edge of the middle turbinate. When the polyp grade score is 3 or 4, the single-frame image includes the lower edge of the inferior turbinate, the common nasal meatus, and the nasal floor.
[0007] Furthermore, the preprocessing of the nasal endoscopic medical image data also includes: The anatomical structures corresponding to the single-frame image are selected and labeled with a grade rating. The grading system adopts the international standard for nasal polyp grading: 0 points for no polyps; 1 point for small polyps in the middle meatus that do not exceed the lower edge of the middle turbinate; 2 points for polyps in the middle meatus that exceed the lower edge of the middle turbinate; 3 points for polyps in the middle meatus that reach the lower edge of the inferior turbinate or for polyps of 2 points that are accompanied by large polyps on the medial side of the middle turbinate that exceed the lower edge of the middle turbinate; 4 points for polyps that reach the nasal floor; the images to be analyzed for polyps of 3 points are divided into subgroups 3a and 3b, where 3a is polyps that reach or exceed the lower edge of the inferior turbinate, and 3b is polyps in both the middle meatus and olfactory cleft that exceed the lower edge of the middle turbinate.
[0008] Furthermore, the data augmentation processing of the training set includes: Obtain the number of samples for each level rating in the training set; If the number of samples is less than a first preset value, the samples with the corresponding level scores are enhanced based on the first enhancement strategy. If the number of samples is less than the second preset value, the samples with the corresponding level scores are enhanced based on the second enhancement strategy. The enhancement strategy includes at least one of angle adjustment, brightness adjustment, and mirroring. The adjustment strategy also includes the number of adjustments. The adjustment range of the first enhancement strategy is greater than the adjustment range of the second enhancement strategy, and the first preset value is less than the second preset value.
[0009] Furthermore, the performance evaluation of the target detection model to be used based on the test set includes: Calculate the mAP50 and F1 scores of each of the target detection models to be used on the validation set; The optimal target detection model to be used is selected based on mAP50 and F1 score.
[0010] Furthermore, after obtaining the automatic grading model for nasal polyps, the process also includes: The confidence threshold of the automatic nasal polyp grading model was adjusted to the highest F1 score to balance accuracy and recall. The highest F1 score was determined by the F1 score calculated by the target detection model to be used on the validation set.
[0011] A scoring method based on an automatic grading model for nasal polyps, comprising: Obtain endoscopic medical images of the nose to be scored; The nasal endoscopic medical image is input into an automatic nasal polyp grading model obtained by any of the above-described automatic nasal polyp grading model training methods, and the nasal polyp grade score corresponding to the nasal endoscopic medical image is output.
[0012] An automatic grading model training system for nasal polyps, the system employing any of the above-described automatic grading model training methods for nasal polyps, specifically comprising the following modules: The first acquisition module is used to acquire endoscopic medical image data of multiple patients with chronic sinusitis. The processing module, connected to the first acquisition module, is used to preprocess the nasal endoscopy medical image data to obtain an endoscopic image set containing nasal polyp scoring and labeling information of different grades. A partitioning module, connected to the processing module, is used to partition the endoscopic image set into a training set, a validation set, and a test set. An enhancement module is used to connect to the partitioning module and to perform data augmentation processing on the training set. The data-augmented training set is used to perform cross-validation training on a preset target detection model to obtain multiple target detection models to be used. An evaluation module, connected to the partitioning module and the enhancement module, is used to perform performance evaluation on the target detection model to be used based on the test set, and select a target detection model to be used based on the performance evaluation result. The optimization module, connected to the partitioning module and the evaluation module, is used to train and optimize the selected target detection model based on the validation set and the data-enhanced training set to obtain an automatic nasal polyp grading model.
[0013] A scoring system based on an automatic grading model for nasal polyps, wherein the system employs a scoring method based on an automatic grading model for nasal polyps as described above, specifically including the following modules: The second acquisition module is used to acquire nasal endoscopic medical images to be scored; The input module, connected to the second acquisition module, is used to input the nasal endoscopic medical image into the nasal polyp automatic grading model obtained by the nasal polyp automatic grading model training method described in any one of the above-mentioned methods, and output the nasal polyp grade score corresponding to the nasal endoscopic medical image.
[0014] An electronic device, the device comprising: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements either the automatic grading model training method for nasal polyps described above or the scoring method based on the automatic grading model for nasal polyps described above.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention constructs an endoscopic image set containing scoring information for different grades of nasal polyps using endoscopic medical imaging data from patients with chronic sinusitis. This image set is then used to train a pre-defined target detection model, enabling the model to directly learn and internalize international standards for nasal polyp scoring. Through a full-process training process involving data augmentation, cross-validation, and optimization, the pre-defined target detection model is ensured to stably learn the characteristics of different grades of nasal polyps. The output nasal polyp scoring is both automatic and accurate, fundamentally overcoming the limitation of artificial intelligence technology, which can only identify polyps but cannot automatically score them. Secondly, data augmentation techniques are used to balance the sample distribution of nasal polyp grade scores in the dataset, solving the problem of training bias in the preset target detection model caused by insufficient samples of some nasal polyp grade scores. The first augmentation strategy simulates extreme variations such as strong light and large angle changes for a few categories. The second augmentation strategy retains the original features of most nasal polyp grade scores while introducing slight perturbations, thereby improving the robustness of the preset target detection model to differences in nasal polyp morphology and location. Attached Figure Description
[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] Figure 1 This is a flowchart illustrating an automatic grading model training method for nasal polyps in Example 1; Figure 2 The 7-fold cross-validation model is illustrated in Example 1; Figure 3This is a flowchart illustrating a scoring method based on an automatic grading model for nasal polyps in Example 2; Figure 4 This is a structural block diagram illustrating an automatic grading model training system for nasal polyps in Example 3; Figure 5 This is a structural block diagram illustrating a scoring system based on an automatic grading model for nasal polyps, as shown in Example 4. Figure 6 This is a structural block diagram illustrating an electronic device in Example 5. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0019] The specific embodiments of the present invention will be described below.
[0020] While existing artificial intelligence technologies attempt to identify nasal polyps, most are limited in function. Some can only determine the presence or absence of polyps, while others focus on distinguishing between polyps and tumors. However, none can automatically grade and score nasal polyps based on their position and size relative to key anatomical structures such as the middle and inferior turbinates. This invention constructs an endoscopic image set containing nasal polyp grading annotations at different levels based on the patient's nasal endoscopic imaging data. This image set is used to train a pre-defined target detection model, enabling the model to directly learn and internalize international standards for nasal polyp grading. Through a full-process training process involving data augmentation, cross-validation, and hyperparameter tuning, the pre-defined target detection model can stably learn nasal endoscopic features covering each grading level. The output nasal polyp grading score is both automatic and accurate, thus solving the core technical bottleneck of automatic endoscopic nasal polyp grading.
[0021] Example 1 like Figure 1 As shown, this invention proposes an automatic grading model training method for nasal polyps, which specifically includes the following steps: Step S11: Obtain endoscopic medical imaging data of the nose from multiple patients with chronic sinusitis.
[0022] In this embodiment, nasal endoscopic medical image data of patients with chronic sinusitis is collected through a hospital endoscopy center or a public database. The nasal endoscopic medical image data is video image. The electronic device can directly connect to the hospital endoscopy center or a public database to automatically obtain the nasal endoscopic medical image data, or it can be manually collected from the hospital endoscopy center or a public database. There are no specific limitations on this.
[0023] Step S12: Preprocess the nasal endoscopic medical image data to obtain an endoscopic image set containing nasal polyp scoring and labeling information of different grades.
[0024] The preprocessing of the nasal endoscopic medical image data includes single-frame image cropping and labeling of nasal polyp grade scores. The single-frame image cropping and nasal polyp grade scoring are explained below.
[0025] For anatomical structure extraction: the nasal endoscopic medical image data is video image, and the preprocessing of the nasal endoscopic medical image data includes: extracting a single frame image from the video image for nasal polyp grading, the single frame image containing the anatomical structures required for grading; when the polyp grading score is 0, 1, or 2, the single frame image includes the ostiomeatal complex structure, the middle nasal meatus, and the lower edge of the middle turbinate; when the polyp grading score is 3 or 4, the single frame image includes the lower edge of the inferior turbinate, the common nasal meatus, and the nasal floor.
[0026] In this embodiment, keyframes are extracted using video editing software, such as Premiere Pro. The keyframes are image frames containing the anatomical structures corresponding to the respective score levels. Specifically, the nasal polyp is placed in the center of the image, occupying approximately 1 / 4 of the endoscopic image size, and includes the following surrounding anatomical structures: For polyps scoring 0, 1, or 2, the image needs to include the complete ostiolus complex structure, focusing on the middle nasal meatus and the lower edge of the middle turbinate; if olfactory cleft polyps are present, they must also be fully exposed. For polyps scoring 3 or 4, the image needs to include the complete lower edge of the inferior turbinate, the common nasal meatus, and the nasal floor. The processed images are then uniformly scaled to a preset size to meet the input requirements of the preset target detection model, where the preset size can be 640×640 pixels.
[0027] For the grading labeling: the anatomical structures corresponding to the single frame image are selected and graded; the grading labeling adopts the international standard for nasal polyp grading, with 0 points for no polyps; 1 point for small polyps in the middle nasal meatus and whose size does not exceed the lower edge of the middle turbinate; 2 points for polyps in the middle nasal meatus and whose size exceeds the lower edge of the middle turbinate; 3 points for polyps in the middle nasal meatus reaching the lower edge of the inferior turbinate or polyps of 2 points accompanied by large polyps on the medial side of the middle turbinate, exceeding the lower edge of the middle turbinate; 4 points for polyps reaching the nasal floor; the images to be analyzed with polyps of 3 points are divided into subgroups 3a and 3b, where 3a is polyps reaching or exceeding the lower edge of the inferior turbinate, and 3b is polyps in the middle nasal meatus and olfactory cleft that both exceed the lower edge of the middle turbinate.
[0028] In this embodiment, each single-frame image is labeled according to the nasal polyp grade score. According to the grade score standard, there are two different judgment criteria for the 3-point image. In order to improve the ability of the preset target detection model to understand the correspondence between polyp location and grade, the 3-point image is further subdivided into 3a and 3b subgroups.
[0029] In this embodiment, professional doctors grade and label the captured single-frame images according to the nasal polyp grading standard. Labeling software (such as LabelImg) is used to select the area including anatomical structures such as nasal polyps, middle nasal meatus, olfactory cleft, middle turbinate, medial olfactory cleft area of middle nasal meatus, inferior turbinate, common nasal meatus to nasal floor, etc., and add grade rating labels corresponding to 0 to 4 points. The 3-point images are divided into 3a and 3b subgroups. The labeled data is cross-validated by multiple doctors to ensure data consistency.
[0030] For the 0-point image, select the middle nasal meatus and middle turbinate; for the 1-point, 2-point, and 3b images, select the nasal polyp, the lower edge of the middle turbinate, and the upper edge of the inferior turbinate with the lower edge of the nasal polyp as the center; for the 3a and 4-point images, select the nasal polyp, the lower edge of the inferior turbinate, and the nasal floor with the lower edge of the nasal polyp as the center.
[0031] In this embodiment, multiple labeled images are combined into an endoscopic image set.
[0032] Step S13: Divide the endoscopic image set into a training set, a validation set, and a test set.
[0033] In this embodiment, the endoscopic image set is randomly divided into a training set, a validation set, and a test set according to the rating in a ratio of 6:2:2. For example, for each rating level, 60% of the images are used for training, 20% for validation, and 20% for testing, and finally merged into a whole dataset.
[0034] Step S14: Perform data augmentation on the training set, and use the data-augmented training set to perform cross-validation training on the preset target detection model to obtain multiple target detection models to be used.
[0035] The data augmentation process specifically includes: obtaining the number of samples for each level score in the training set; when the number of samples is less than a first preset value, performing augmentation processing on the samples for the corresponding level score based on a first augmentation strategy; when the number of samples is less than a second preset value, performing augmentation processing on the samples for the corresponding level score based on a second augmentation strategy; wherein the augmentation strategy includes at least one of angle adjustment, brightness adjustment, and mirroring, and the adjustment strategy also includes the number of adjustments, the adjustment range of the first augmentation strategy is greater than the adjustment range of the second augmentation strategy, and the first preset value is less than the second preset value.
[0036] For example, for a grade rating with a sample size less than a first preset value, such as subgroups 3a and 3b after dividing a 3-point image, a first enhancement strategy is used to generate new samples multiple times. The first enhancement strategy includes, but is not limited to, randomly rotating the original image by 90 degrees, adjusting the brightness by 50%, and mirroring the image. For a grade rating with a sample size less than a second preset value, such as 0 or 1 point, a second enhancement strategy is used to generate new samples a few times. The first enhancement strategy includes, but is not limited to, randomly rotating the original image by 10 degrees, adjusting the brightness by 10%, and mirroring the image.
[0037] In this embodiment, data augmentation techniques are used to balance the sample distribution of each level of score in the training set, thereby solving the problem of training bias in the preset target detection model caused by insufficient samples of some nasal polyp level scores. The first augmentation strategy simulates extreme variations such as strong light and large angle changes for a few categories. The second augmentation strategy retains the original features of most level scores while introducing slight perturbations, thereby improving the robustness of the preset target detection model to differences in the morphology and location of nasal polyps.
[0038] The mirroring of subgroups 3a and 3b enhances the understanding of the spatial symmetry of polyps by the preset target detection model, avoids the preset target detection model from overfitting to the anatomical structure of a specific direction of a nostril, and thus improves the accuracy of nasal polyp grade scoring.
[0039] In this embodiment, the YOLOv11n model is used as the preset target detection model, and cross-validation is performed as follows: Figure 2 The 7-fold cross-validation model shown reduces evaluation bias caused by random data splitting by repeatedly dividing the training and validation subsets, ensuring that the preset target detection model performs consistently on unseen data.
[0040] In this embodiment, a 7-fold cross-validation model is used to train a preset target detection model, thereby obtaining multiple target detection models to be used.
[0041] Step S15: Perform performance evaluation on the target detection model to be used based on the test set, and select a target detection model to be used based on the performance evaluation results.
[0042] Specifically, the mAP50 and F1 scores of each target detection model to be used are calculated on the validation set; an optimal target detection model to be used is selected based on the mAP50 and F1 scores.
[0043] In this embodiment, the performance metrics of each target detection model to be used on the validation set are calculated. The performance metrics include mAP50 and F1 score. mAP50 comprehensively measures the target detection and classification capabilities of the target detection model under relaxed localization requirements, while F1 score focuses on evaluating the balance of the classification results of the target detection model to be used.
[0044] In this embodiment, based on the above performance indicators, a comprehensive evaluation strategy is adopted to select the optimal model. For example, weights can be assigned to mAP50 and F1 score for weighted summation, or the model with the highest mAP50 can be selected first, and the F1 score can be compared when they are comparable. By combining these two indicators, the classification ability of the model can be comprehensively evaluated, effectively avoiding the problem of overfitting or underfitting the model in specific level scores (such as 3a and 3b or 2 points and 3b level, etc., with ambiguous boundaries). Thus, the model configuration with the strongest generalization ability and the highest robustness in the automatic nasal polyp grading task is selected.
[0045] Step S16: Based on the validation set and the data-enhanced training set, train and hyperparameter-tuning the selected target detection model to be used to obtain the automatic nasal polyp grading model.
[0046] In this embodiment, all the enhanced training sets and the original validation sets are combined as training data.
[0047] Training period: set to 300 epochs; Learning rate: adopts a linear decay strategy, with the initial learning rate set to 0.001 and the final learning rate decaying to 0.0001; Batch size: set to 16.
[0048] In contrast, the hyperparameters used in the aforementioned cross-validation training phase were: Training period: 1000 epochs; learning rate: kept constant at 0.01; batch size: set to 32.
[0049] Through the aforementioned differentiated hyperparameter settings, the cross-validation stage focuses on fully exploring the model's potential with a larger learning rate, while step S16 focuses on fine-tuning the selected optimal model with a smaller learning rate, thereby obtaining the best-performing automatic nasal polyp grading model.
[0050] By training and hyperparameter tuning the selected target detection model, the problem of local overfitting or underfitting in cross-validation can be reduced. Hyperparameter optimization can balance the convergence speed and stability of the automatic nasal polyp grading model and avoid training oscillations. Training the target detection model on the validation set can enhance the sensitivity of the automatic nasal polyp grading model to the subdivision features of nasal polyps and improve its discrimination ability on boundary cases, especially the ambiguous samples between 3b and 2.
[0051] After obtaining the automatic grading model for nasal polyps, the method further includes: adjusting the confidence threshold of the automatic grading model for nasal polyps to the highest point of the F1 score to balance accuracy and recall, wherein the highest point of the F1 score is determined by the F1 score calculated by the target detection model to be used on the validation set.
[0052] In this embodiment, the trained automatic nasal polyp grading model is deployed to a computer or a software program embedded in a nasal endoscope. Based on the results of the validation set, the confidence threshold of the automatic nasal polyp grading model is adjusted to the highest point of the F1 score to balance accuracy and recall.
[0053] Example 2 like Figure 3 As shown, this invention proposes a scoring method based on an automatic grading model for nasal polyps, which specifically includes the following steps: Step S21: Obtain the nasal endoscopic medical images to be scored.
[0054] Step S22: Input the nasal endoscopic medical image into the nasal polyp automatic grading model obtained by the nasal polyp automatic grading model training method described in Example 1, and output the nasal polyp grade score corresponding to the nasal endoscopic medical image.
[0055] In this embodiment, the nasal endoscopic medical images that need to be scored are input into the trained automatic nasal polyp grading model, and the nasal polyp grade score corresponding to the nasal endoscopic medical images can be output. The nasal endoscopic medical images are single-frame images.
[0056] Example 3 like Figure 4 As shown, the present invention also proposes an automatic grading model training system for nasal polyps, using an automatic grading model training method for nasal polyps as described in Example 1, comprising the following modules: The first acquisition module is used to acquire endoscopic medical image data of multiple patients with chronic sinusitis. The processing module, connected to the first acquisition module, is used to preprocess the nasal endoscopy medical image data to obtain an endoscopic image set containing nasal polyp scoring and labeling information of different grades. A partitioning module, connected to the processing module, is used to partition the endoscopic image set into a training set, a validation set, and a test set. An enhancement module is used to connect to the partitioning module and to perform data augmentation processing on the training set. The data-augmented training set is used to perform cross-validation training on a preset target detection model to obtain multiple target detection models to be used. An evaluation module, connected to the partitioning module and the enhancement module, is used to perform performance evaluation on the target detection model to be used based on the test set, and select a target detection model to be used based on the performance evaluation result. The optimization module, connected to the partitioning module and the evaluation module, is used to train and optimize the selected target detection model based on the validation set and the data-enhanced training set to obtain an automatic nasal polyp grading model.
[0057] This invention addresses the limitation of existing artificial intelligence technologies in automatically grading nasal polyps based on their position and size relative to key anatomical structures such as the middle and inferior turbinates. Based on patients' nasal endoscopic imaging data, this invention constructs an endoscopic image set containing nasal polyp grading annotations for different grades. This image set is then used to train a pre-defined target detection model, enabling the model to directly learn and internalize international standards for nasal polyp grading. Furthermore, through a full-process training process involving data augmentation, cross-validation, and hyperparameter tuning, the pre-defined target detection model stably learns nasal endoscopic features covering each grading level. The output nasal polyp grading score is both automatic and accurate, thus overcoming the core technical bottleneck of automatic endoscopic nasal polyp grading.
[0058] Example 4 like Figure 5 As shown, this invention also proposes a scoring system based on an automatic grading model for nasal polyps, using a scoring method based on an automatic grading model for nasal polyps as described in Example 2, including the following modules: The second acquisition module is used to acquire nasal endoscopic medical images to be scored; The input module, connected to the second acquisition module, is used to input the nasal endoscopic medical image into the nasal polyp automatic grading model obtained by the nasal polyp automatic grading model training method described in Example 1, and output the nasal polyp grade score corresponding to the nasal endoscopic medical image.
[0059] To address the problem that existing artificial intelligence technologies cannot grade nasal polyps based on their position and size relative to key anatomical structures such as the middle and inferior turbinates, this invention inputs endoscopic medical images into a trained automatic nasal polyp grading model, which can directly obtain the corresponding nasal polyp grade score.
[0060] Example 5 Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device includes a memory, a processor, and a communication bus; the memory and the processor are connected via the communication bus. The memory stores a method for training an automatic nasal polyp grading model or a scoring method based on an automatic nasal polyp grading model, which can be loaded and executed by the processor, as provided in the above embodiments.
[0061] The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function, and instructions for implementing the automatic nasal polyp grading model training method or the scoring method based on the automatic nasal polyp grading model provided in the above embodiments. The data storage area may store data involved in the automatic nasal polyp grading model training method or the scoring method based on the automatic nasal polyp grading model provided in the above embodiments.
[0062] The processor may include one or more processing cores. The processor executes instructions, programs, code sets, or instruction sets stored in memory, and calls data stored in memory to perform various functions and process data as described in this application. The processor may be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, and microprocessor. It is understood that, for different devices, the electronic devices used to implement the above-described processor functions may also be other types, and the embodiments of this application do not specifically limit this.
[0063] A communication bus can include a pathway for transmitting information between the aforementioned components. The communication bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Communication buses can be categorized into address buses, data buses, control buses, etc. For ease of representation, Figure 6 The symbol is represented by a single double arrow, but this does not mean that there is only one bus or one type of bus.
[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.
Claims
1. A method for training an automatic grading model for nasal polyps, characterized in that, include: Acquire endoscopic medical imaging data of the nose from multiple patients with chronic sinusitis; The nasal endoscopic medical image data is preprocessed to obtain an endoscopic image set containing scoring and labeling information for different grades of nasal polyps; The endoscopic image set is divided into a training set, a validation set, and a test set; The training set is subjected to data augmentation processing, and the pre-set target detection model is cross-validated and trained using the data augmented training set to obtain multiple target detection models to be used. The performance of the target detection model to be used is evaluated based on the test set, and a target detection model to be used is selected based on the performance evaluation results. Based on the validation set and the data-enhanced training set, the selected target detection model is trained and its hyperparameters are tuned to obtain an automatic nasal polyp grading model.
2. The method for training an automatic grading model of nasal polyps according to claim 1, characterized in that, The nasal endoscopic medical imaging data is video image, and the preprocessing of the nasal endoscopic medical imaging data includes: A single frame image for nasal polyp grading is extracted from the video image, the single frame image containing the anatomical structures required for grading; When the polyp grade score is 0, 1 or 2, the single frame image includes the ostiomeatal complex structure, the middle meatus and the lower edge of the middle turbinate. When the polyp grade score is 3 or 4, the single-frame image includes the lower edge of the inferior turbinate, the common nasal meatus, and the nasal floor.
3. The method for training an automatic grading model of nasal polyps according to claim 2, characterized in that, The preprocessing of the nasal endoscopic medical image data further includes: The anatomical structures corresponding to the single-frame image are selected and labeled with a grade rating. The grading system adopts the international standard for nasal polyp grading: 0 points for no polyps; 1 point for small polyps in the middle meatus that do not exceed the lower edge of the middle turbinate; 2 points for polyps in the middle meatus that exceed the lower edge of the middle turbinate; 3 points for polyps in the middle meatus that reach the lower edge of the inferior turbinate or for polyps of 2 points that are accompanied by large polyps on the medial side of the middle turbinate that exceed the lower edge of the middle turbinate; 4 points for polyps that reach the nasal floor; the images to be analyzed for polyps of 3 points are divided into subgroups 3a and 3b, where 3a is polyps that reach or exceed the lower edge of the inferior turbinate, and 3b is polyps in both the middle meatus and olfactory cleft that exceed the lower edge of the middle turbinate.
4. The method for training an automatic grading model of nasal polyps according to claim 1, characterized in that, The data augmentation process performed on the training set includes: Obtain the number of samples for each level rating in the training set; If the number of samples is less than a first preset value, the samples with the corresponding level scores are enhanced based on the first enhancement strategy. If the number of samples is less than the second preset value, the samples with the corresponding level scores are enhanced based on the second enhancement strategy. The enhancement strategy includes at least one of angle adjustment, brightness adjustment, and mirroring. The adjustment strategy also includes the number of adjustments. The adjustment range of the first enhancement strategy is greater than the adjustment range of the second enhancement strategy, and the first preset value is less than the second preset value.
5. The method for training an automatic grading model of nasal polyps according to claim 1, characterized in that, The performance evaluation of the target detection model to be used based on the test set includes: Calculate the mAP50 and F1 scores of each of the target detection models to be used on the validation set; The optimal target detection model to be used is selected based on mAP50 and F1 score.
6. A scoring method based on an automatic grading model for nasal polyps, characterized in that, include: Obtain endoscopic medical images of the nose to be scored; The nasal endoscopic medical image is input into the nasal polyp automatic grading model obtained by the nasal polyp automatic grading model training method according to any one of claims 1-5, and the nasal polyp grade score corresponding to the nasal endoscopic medical image is output.
7. An automatic grading model training system for nasal polyps, characterized in that, The system employs an automatic nasal polyp grading model training method according to any one of claims 1 to 5, specifically including the following modules: The first acquisition module is used to acquire endoscopic medical image data of multiple patients with chronic sinusitis. The processing module, connected to the first acquisition module, is used to preprocess the nasal endoscopy medical image data to obtain an endoscopic image set containing nasal polyp scoring and labeling information of different grades. A partitioning module, connected to the processing module, is used to partition the endoscopic image set into a training set, a validation set, and a test set. An enhancement module is used to connect to the partitioning module and to perform data augmentation processing on the training set. The data-augmented training set is used to perform cross-validation training on a preset target detection model to obtain multiple target detection models to be used. An evaluation module, connected to the partitioning module and the enhancement module, is used to perform performance evaluation on the target detection model to be used based on the test set, and select a target detection model to be used based on the performance evaluation result. The optimization module, connected to the partitioning module and the evaluation module, is used to train and optimize the selected target detection model based on the validation set and the data-enhanced training set to obtain an automatic nasal polyp grading model.
8. A scoring system based on an automatic grading model for nasal polyps, characterized in that, The system employs a scoring method based on an automatic grading model for nasal polyps as described in claim 6, specifically including the following modules: The second acquisition module is used to acquire nasal endoscopic medical images to be scored; The input module, connected to the second acquisition module, is used to input the nasal endoscopic medical image into the nasal polyp automatic grading model obtained by the nasal polyp automatic grading model training method according to any one of claims 1-5, and output the nasal polyp grade score corresponding to the nasal endoscopic medical image.
9. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements a method for training an automatic grading model of nasal polyps as described in any one of claims 1-5 or a scoring method based on an automatic grading model of nasal polyps as described in claim 6.