A Deep Learning-Based Method and System for Automatic Assessment of Adenoid Hypertrophy in Children Using X-ray Images
By constructing a deep learning model based on ResNet-50 and Keypoint R-CNN, the problems of large error and insufficient sample in the diagnosis of adenoid hypertrophy in X-ray evaluation were solved, and highly accurate automatic assessment of adenoid hypertrophy was achieved, reducing the workload of clinical practice.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI TONGJI STOMATOLOGY HOSPITAL (TONGJI UNIVERSITY AFFILIATED STOMATOLOGY HOSPITAL)
- Filing Date
- 2026-02-10
- Publication Date
- 2026-05-26
AI Technical Summary
Existing diagnostic models for adenoid hypertrophy suffer from large errors and insufficient sample size in X-ray evaluation, leading to inaccurate diagnostic results.
We adopted a deep learning-based approach, using ResNet-50 and Keypoint R-CNN architectures to construct an intelligent assessment model for adenoid hypertrophy in children through data annotation, standardization, and data augmentation. We then used the A/N ratio method to automatically assess adenoid hypertrophy.
It improved the accuracy and reliability of adenoid hypertrophy diagnosis, reduced the workload of clinicians, and mitigated the impact of annotation errors, achieving an accuracy of 93.83%, a sensitivity of 87.10%, and a specificity of 95.42%.
Smart Images
Figure CN122089686A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent identification system for adenoid assessment, specifically involving a method and system for automatically assessing adenoid hypertrophy in children based on X-ray images using deep learning. Background Technology
[0002] Adenoid and tonsil hypertrophy is the most common cause of mouth breathing in children. Pathologically enlarged adenoids reduce the cross-sectional area of the pharynx and obstruct nasal breathing, forcing children to breathe entirely or partially through their mouths to obtain sufficient oxygen. The function and morphology of the orofacial system are unified; the habitual position of muscles inside and outside the mouth affects tooth development. In patients with mouth breathing, the physiological breathing pattern is disrupted, and abnormal muscle habits and positions lead to changes in the oral cavity and craniofacial region. Although the most accurate diagnostic method for adenoid hypertrophy is currently endoscopy, in clinical practice, some children often cannot cooperate well during endoscopic treatment, increasing the difficulty of diagnosis. In orthodontic clinical practice, cephalometric radiography is a simple, economical, and routine examination. The method for diagnosing adenoid hypertrophy (AH) based on cephalometric radiography, proposed by Fujioka et al. in 1979, utilizes the A / N ratio. This involves measuring the thickness (A) of the adenoids and the width (N) of the nasopharynx on a cephalometric radiograph. A / N ≤ 0.60 indicates adenoid hypertrophy within the normal range; 0.61 ≤ A / N ≤ 0.70 indicates moderate adenoid hypertrophy; A / N ≥ 0.71 indicates pathological adenoid hypertrophy; and A / N ≥ 0.80 indicates significant adenoid hypertrophy. However, the entire assessment process is complex and time-consuming, and inaccurate identification of cephalographic markers can lead to incorrect assessment results. Therefore, developing an algorithm to automatically classify AH in cephalometric radiographs is necessary. Currently, research on deep learning for AH assessment in lateral cephalometric radiographs is still limited.
[0003] Current research on the application of deep learning in the assessment of adenoid hypertrophy (AH) on lateral cephalometric radiographs remains limited. Existing technologies, such as CN116630324B, disclose a method for automatically assessing adenoid hypertrophy based on deep learning in MRI images. However, this method processes MRI images, which have high soft tissue resolution, while lateral cephalometric radiographs are X-ray projection images, which suffer from structural overlap and low grayscale contrast, requiring targeted preprocessing and model design.
[0004] The closest existing model to our model is HeadNet. Its training and design process involved collecting 901 lateral cephalometric radiographs, randomly selecting 581 for training, 160 for validation, and the remaining 160 for testing. Four points were labeled in the training set for the diagnosis of adenoid hypertrophy (Ba, Ar, A(T), PNS). The model also utilized methods such as translation and rotation to expand the dataset and improve the generalization ability of the neural network. However, the choice of HeadNet as the model has the following drawbacks and limitations:
[0005] 1. When diagnosing adenoid hypertrophy, this model uses the A / N ratio anterior wall method. However, in its application, it treats the intersection of the distances between A(T) and the lines connecting Ba and Ar as the juncture point, which has a certain positional difference from the actual occipital-sphenoid junction point. This may cause a certain degree of error in the final diagnosis.
[0006] 2. The initial sample size of the model was small. Although the sample size was increased by methods such as rotation and translation, it still affected the applicability of the model to some extent. Summary of the Invention
[0007] The technical problem to be solved by the present invention is to provide a method and system for automatically assessing adenoid hypertrophy in children based on X-ray images using deep learning, which solves the problem of large error in the discrimination results output by the existing adenoid hypertrophy diagnostic model.
[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0009] A method for automatically assessing adenoid hypertrophy in children based on X-ray images using deep learning, comprising the following steps:
[0010] Step S1: Divide the acquired X-ray image data into three parts: training set, validation set, and test set.
[0011] Step S2: Label the image data in the training and validation sets;
[0012] Step S3: Standardize the image data labeled in the training and validation sets;
[0013] Step S4: Construct an intelligent assessment model for adenoid hypertrophy in children based on ResNet-50 convolutional neural network;
[0014] Step S5: Use the training set and validation set obtained in the previous step to train the evaluation model and generate an intelligent training model for adenoid hypertrophy in children.
[0015] Step S6: Evaluate the training model generated in step S5 using the test set obtained in step S1.
[0016] In step S2, the specific operations for data annotation are as follows:
[0017] The labelme tool was used to mark five anatomical landmarks required for adenoid hypertrophy identification: the posterior border of the hard palate, the lower border of the clivus, the most prominent point of the adenoid, the junction of the pterygoid plate and the skull base, and the intersection of the surface of the lower skull base and the average posterior surface of the mandibular condyle. The image data was then converted from JPG to JSON format.
[0018] The specific operation of step S3 is as follows:
[0019] Perform flip, rotate, translate, and scale operations on JSON formatted X-ray images, while simultaneously adjusting the coordinates of key points and increasing the number of samples.
[0020] Step S3 also includes a preprocessing sub-step, which performs noise reduction, smoothing, and sharpening on the image to enhance image contrast, and then normalizes the pixel values to the [0-1] range through grayscale normalization. Step S4, which constructs an intelligent assessment model for pediatric adenoid hypertrophy, includes:
[0021] The feature extraction unit extracts multi-dimensional features of the image by stacking residual blocks of ResNet-50 and outputs feature maps at different levels.
[0022] Anchor frame generation unit: AnchorGenerator is used to generate anchor frames with sizes ranging from 32 to 512 pixels and aspect ratios ranging from 0.25 to 4.0, for adenoid region localization.
[0023] The keypoint regression unit, based on the Keypoint R-CNN architecture, predicts the keypoint locations of each target through convolution operations and non-maximum suppression.
[0024] The specific operation method of step S5 is as follows:
[0025] The evaluation model was trained using the training set image data after the standardization process in step S3. The learning rate was set to 0.001, the momentum to 0.9, the weight decay to 0.0005, and the optimizer to be the SGD / Adam optimizer. The model was then validated using the validation set image data after the standardization process in step S3 using cross-validation, thus generating an intelligent training model for pediatric adenoid hypertrophy.
[0026] The specific operation method of S6 is as follows:
[0027] Using the test set obtained in step S1, the trained model is tested, and the A / N value is calculated using the five predicted key points to determine whether it is adenoid hypertrophy.
[0028] A deep learning-based system for automatically assessing pediatric adenoid hypertrophy using X-ray images includes a data processing and loading module. This module loads X-ray images (JPG) and key point annotations (CSV / JSON), performs standardization, centers the images to 1200*1500 pixels, normalizes them according to ImageNet parameters, and converts them to CHW tensors. Finally, it performs data augmentation through flipping, rotation, translation, scaling, and other techniques.
[0029] The core architecture module of the model constructs a network architecture for keypoint detection, enabling feature extraction, candidate box generation, and keypoint regression.
[0030] The training and optimization module is responsible for the model's training loop, parameter updates, and learning rate scheduling, ensuring model convergence and optimizing detection accuracy.
[0031] The prediction post-processing module filters and corrects the model's original prediction results and outputs accurate key point coordinates.
[0032] The output module inputs real-time acquired X-ray image data into the automatic assessment model for adenoid hypertrophy in children, and automatically obtains the assessment level of adenoid hypertrophy in children.
[0033] It also includes a 3D visualization report model, which automatically generates and outputs a visualized diagnostic report containing key point predictions, grading conclusions, and image annotations.
[0034] A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, invoke the steps of the method.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] 1. This solution utilizes an algorithm processing module to process acquired data such as lateral cephalometric radiographs, and identifies key points in the image data using the Keypoint R-CNN_ResNet-50_FPN model. It also employs the A / N value anterior wall method to construct adenoidectic diagnostic grading and outputs analysis results. This solution addresses the issue of insufficient accuracy when MRI image processing methods are directly applied to X-ray images.
[0037] 2. A novel automatic identification model for adenoid hypertrophy based on deep learning was developed. The model demonstrated good accuracy and reliability in adenoid hypertrophy identification (accuracy 93.83%, sensitivity 87.10%, specificity 95.42%), and reduced the complexity of adenoid hypertrophy identification in children, effectively alleviating the workload of clinicians.
[0038] 3. Compared with the HeadNet model, the ResNet-50 model in this solution has a better performance in key point recognition, and the smooth L1 loss function reduces the impact of annotation error, further improving the recognition accuracy. Attached Figure Description
[0039] Figure 1 This describes the key point annotation process for the lateral cephalometric radiograph of the adenoid hypertrophy identification model of this invention.
[0040] Figure 2 This is the result of data enhancement for lateral cephalometric radiographs of children according to the present invention.
[0041] Figure 3 The key point prediction results of the adenoid hypertrophy identification model for children in this invention are shown (red points are manually marked points, and green points are predicted points by the model).
[0042] Figure 4 The results are the identification results of the adenoid hypertrophy identification model for children according to the present invention. Detailed Implementation
[0043] The structure and working process of the present invention will be further described below with reference to the accompanying drawings.
[0044] The purpose of this invention is to address the problems in the background technology, namely that lateral cephalometric radiographs (X-ray images) have significant differences in imaging principles and image features (such as contrast, noise distribution, and clarity of anatomical structure display): MRI images have high soft tissue resolution, while lateral cephalometric radiographs are X-ray projection imaging, which has problems such as structural overlap and low grayscale contrast, requiring targeted preprocessing and model design. This invention proposes an intelligent recognition system and method for adenoid hypertrophy in children.
[0045] A method for automatically assessing adenoid hypertrophy in children based on X-ray images using deep learning, comprising the following steps:
[0046] Step S1: Divide the acquired X-ray image data into three parts: training set, validation set, and test set.
[0047] Step S2: Label the image data in the training and validation sets;
[0048] Step S3: Standardize the image data labeled in the training and validation sets;
[0049] Step S4: Construct an intelligent assessment model for adenoid hypertrophy in children based on ResNet-50 convolutional neural network;
[0050] Step S5: Use the training set and validation set obtained in the previous step to train the evaluation model and generate an intelligent training model for adenoid hypertrophy in children.
[0051] Step S6: Evaluate the training model generated in step S5 using the test set obtained in step S1.
[0052] Furthermore, in step S2, the specific operation of data annotation is as follows:
[0053] Five anatomical landmarks required for adenoid hypertrophy identification were marked by professional medical personnel using the labelme tool: the posterior border of the hard palate (PNS), the lower border of the clivus (Ba), the most prominent point of the adenoid (T), the junction of the pterygoid plate and the skull base (W), and the intersection of the surface of the lower skull base and the average posterior surface of the mandibular condyle (Ar). The image data was also converted from JPG to JSON format.
[0054] Furthermore, in step S3, the specific operation of standardizing the training set and validation set is as follows:
[0055] Image data is center-cropped through a data preprocessing channel, and the image size is fixed at 1200*1500Px; data augmentation is performed by flipping, rotating, translating, scaling and other methods to increase the number of samples; grayscale normalization is performed to unify pixel values to the [0-1] range.
[0056] Furthermore, in step S4, the ResNet-50 pediatric adenoid hypertrophy assessment model based on RCNN includes:
[0057] Feature extraction unit: ResNet-50 is used as the backbone network. A deep network structure is built by stacking residual blocks to avoid gradient vanishing or exploding problems. It can learn complex feature representations of images and output feature maps at different levels (shallow features contain spatial details, and deep features contain semantic information).
[0058] Feature Fusion Unit: The multi-scale feature maps output by ResNet-50 are fused through FPN (Feature Pyramid Network). By combining the semantic information of high-level features with the spatial details of low-level features through top-down upsampling paths and lateral connections, a multi-scale feature pyramid is generated, which significantly improves the detection performance of small targets (such as adenoid key points).
[0059] Anchor frame generation unit: The AnchorGenerator is used to generate anchor frames of various sizes and aspect ratios, covering sizes from 32 to 512 pixels (32, 64, 128, 256, 512) and aspect ratios from 0.25 to 4.0 (0.25, 0.5, 0.75, 1.0, 2.0, 3.0, 4.0). This allows for adaptation to adenoid regions of different sizes and shapes, improving the accuracy of region localization.
[0060] Keypoint Regression Unit: Based on the Keypoint R-CNN architecture, it generates keypoint heatmaps through a fully convolutional network after region localization. These heatmaps are then decoded into coordinate values through post-processing. Non-maximum suppression (IOU threshold of 0.3) is used to remove redundant predictions, ensuring the accuracy of keypoint predictions. The model training employs a smooth L1 loss function to reduce the impact of annotation errors and improve the model's robustness.
[0061] Furthermore, the specific operation of step S5 is as follows:
[0062] The ResNet-50 was trained using the training set image data that had been standardized in step S3, with a learning rate of 0.001, momentum of 0.9, weight decay of 0.0005, and an SGD / Adam optimizer. The model was then validated using the validation set image data that had been standardized in step S3.
[0063] Furthermore, the specific operation process of step S6 is as follows:
[0064] Using the unprocessed image data test set obtained in step S1, the trained model is tested to predict five landmarks and calculate the A / N value to determine whether it is adenoid hypertrophy.
[0065] Specific embodiments, such as Figures 1 to 4 As shown:
[0066] This example provides a method for building and training a model for automatically identifying adenoid hypertrophy based on X-ray images using deep learning, illustrated below:
[0067] 1. Data collection and filtering
[0068] This method first establishes a standardized medical image acquisition process, obtaining digital lateral cephalometric X-ray images of children aged 3-10 years from specialized hospitals. The sample format is JPG, and the image size is 2050*1687Px. All images must meet the following conditions:
[0069] (1) It contains a complete and clear nasopharyngeal anatomy;
[0070] (2) Approved by the ethics committee;
[0071] Blurry, distorted, or warped image data, as well as images with abnormal craniofacial morphology and structure, were excluded. Finally, 1499 valid images were obtained and divided into training, validation, and test sets.
[0072] 2. Key point annotation
[0073] The raw data was processed using the LabelMe tool, and the labeled data was saved in a CVS file. Each data point (key point + head shadow image) was distinguished by a unique identifier.
[0074] Establish a labeling system comprising five anatomical landmarks, including:
[0075] - Posterior border of the hard palate (PNS)
[0076] - The lowest posterior point of the margin of the foramen magnum (Ba)
[0077] - The most prominent point of the adenoid (T)
[0078] - The intersection of the base of the lower skull and the average posterior surface of the mandibular condyle (Ar)
[0079] - The junction of the occipital and sphenoid bones (W)
[0080] 3. Construction of a hierarchical recognition model
[0081] 3.1 Preprocessing channel.
[0082] The image first passes through a preprocessing pipeline to provide standardized input for the intelligent recognition model:
[0083] Center crop: Fixed crop to 1200×1500 pixels (achieved via imgaug.CropToFixedSize), while simultaneously adjusting the coordinates of 5 key points in the image.
[0084] Data augmentation (training phase): Perform random flipping (horizontal / vertical, probability 0.5), rotation (±30°), translation (±10%), scaling (±20%), Gaussian noise (variance 10-50), and brightness / contrast adjustment (±20%) via alumniation.
[0085] Standardization: Pixel values are normalized to ImageNet mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225].
[0086] Format conversion: Convert the NumPy array (H, W, C) to a PyTorch tensor (C, H, W) with pixel values in the range [0,1].
[0087] 3.2 Building an Intelligent Recognition Model
[0088] Feature extraction unit: ResNet-50 is used as the backbone network. A deep network structure is built by stacking residual blocks to avoid gradient vanishing or exploding problems. It can learn complex feature representations of images and output feature maps at different levels (shallow features contain spatial details, and deep features contain semantic information).
[0089] Feature Fusion Unit: The multi-scale feature maps output by ResNet-50 are fused through FPN (Feature Pyramid Network). By combining the semantic information of high-level features with the spatial details of low-level features through top-down upsampling paths and lateral connections, a multi-scale feature pyramid is generated, which significantly improves the detection performance of small targets (such as adenoid key points).
[0090] Anchor frame generation unit: AnchorGenerator is used to generate anchor frames of various sizes and aspect ratios, covering sizes from 32 to 512 pixels and aspect ratios from 0.25 to 4.0, which can adapt to adenoid regions of different sizes and shapes, improving the accuracy of region positioning;
[0091] Keypoint Regression Unit: Based on the Keypoint R-CNN architecture, it generates keypoint heatmaps through a fully convolutional network after region localization. These heatmaps are then decoded into coordinate values through post-processing. Non-maximum suppression (IOU threshold of 0.3) is used to remove redundant predictions, ensuring the accuracy of keypoint predictions. The model training employs a smooth L1 loss function to reduce the impact of annotation errors and improve the model's robustness.
[0092] 3.3 Training of the Intelligent Recognition Model
[0093] The ResNet-50 was trained using standardized training image data, with the SGD / Adam optimizer selected. The learning rate was set to 0.001, momentum to 0.9, and weight decay to 0.0005. The training epochs were set to 100, with the learning rate decaying to 0.3 times the original value every 5 epochs. After each training epoch, the results were validated using standardized validation image data. Training was stopped when the validation set performance converged.
[0094] 3.4 Identification and Analysis Module
[0095] Quantitative analysis system based on labeled key points:
[0096] (1) Calculate the A / N ratio: The vertical distance between the largest adenoid tissue point (T) and the straight line Ba-Ar (skull base tangent) is the thickness of the adenoid (A), and the distance between the junction of the pterygoid plate and the skull base (W) and the posterosuperior segment of the hard palate (PNS) is the width of the nasopharynx (N).
[0097] (2) Establish a four-level diagnostic standard:
[0098] - Normal (A / N≤0.60)
[0099] - Mild hypertrophy (0.61-0.70)
[0100] - Moderate hypertrophy (0.71-0.80)
[0101] - Severe hypertrophy (≥0.81)
[0102] (3) Develop a three-dimensional visualization report system to automatically generate a visualization diagnostic report containing key point prediction values, hierarchical conclusions and image annotations.
[0103] 3.5 Model Evaluation:
[0104] In this example, the test set is used to evaluate the model's performance. The test set data was never used during training, therefore its evaluation results better reflect the model's performance in real-world applications.
[0105] The specific method is as follows:
[0106] Image data from the test set is input into the trained model. Initial feature extraction is performed using conv1 (7×7 convolutional kernel, stride 2) + BN layer + ReLU layer. The output is then downsampled by a 3×3 max pooling layer (stride 2). Subsequently, it is processed by conv2_x The four residual block groups (containing 16 residual blocks) from conv2_x to conv5_x are generated. Each residual block consists of 1×1, 3×3, and 1×1 convolutional layers connected in series, with BN and ReLU layers interspersed in between, resulting in four sets of feature maps at different scales: C2 (1 / 4 resolution), C3 (1 / 8), C4 (1 / 16), and C5 (1 / 32), achieving feature progression from low to high levels. The FPN module fuses the C2-C5 feature maps output from conv2_x to conv5_x. It unifies the number of channels to 256 dimensions through 1×1 convolution, and then adds them together through bilinear interpolation upsampling and lateral connections to generate feature pyramids at four scales: P2-P5. Finally, the multi-scale fused local features are obtained and sent to the decoder.
[0107] In the decoder, the input multi-scale feature map is first fed into the RPN (Region Proposal Network), which generates candidate boxes based on a preset anchor generator (anchor size 32 / 64 / 128 / 256 / 512, aspect ratio 0.25-4.0). After being filtered by NMS, the fixed-dimensional features of the candidate regions are extracted from the feature pyramid by RoIAlign. Then, it enters the detection head based on adaptive convolution. Each detection head contains two parallel branches: one branch extracts the global features of the target through a 3×3 convolutional layer + BN layer + ReLU layer for classification and bounding box regression; the other branch is a deformable convolutional layer that adaptively captures the flexible deformation features of the target in the X-ray image. The outputs of the two branches are added element-wise and then upsampled by transposed convolution (3×3 convolutional kernel, stride 2, padding 1) to restore the feature map resolution. After iterative optimization by three such sub-modules, the output of the last upsampling layer is a feature map that matches 1 / 4 of the resolution of the original image and is sent to the landmark detection module.
[0108] In the landmark detection module, a key point heatmap is first generated through a 1×1 convolutional layer. After reshaping, the heatmap is fed into the DSNT layer. By performing probabilistic modeling and coordinate regression on the heatmap, the final coordinate values are directly output. Finally, the AN ratio is calculated based on the coordinates.
[0109] The model's test set outputs are compared with the actual test set results to obtain TP (true positive), FP (false positive), TN (true negative), and FN (false negative). Based on this data, relevant parameters are calculated using the following method:
[0110] Acc. = (TP + TN) / (TP + TN + FP + FN)
[0111] Sens. = TP / (TP + FN)
[0112] Spec. = TN / (FP + TN).
[0113] In this example, a data augmentation strategy (flip, rotate ±30°, translate, scale, scale 0.8-1.2 times) was designed to address the low contrast and overlapping structures in X-ray images. Image preprocessing included noise reduction, sharpening, and enhancement of anatomical structure clarity, enabling the model to adapt to image differences caused by different imaging devices, shooting angles, and children's ages. It maintains stable performance in diverse clinical data scenarios without the need for manual image screening or preprocessing.
[0114] In this example, the system features a modular design, automating the entire process of data loading, model training, and predictive inference; it provides visualization tools (key point comparison charts, diagnostic reports) so that doctors can intuitively verify the accuracy of the results; and it supports saving and loading model weights, eliminating the need for repeated training after deployment.
[0115] In this example, the code is compatible with CPU / GPU environments and supports multi-GPU parallel training and inference. The batch size and number of worker processes can be flexibly configured via command line to adapt to different hardware computing power. It can be deployed on local servers in hospitals or ordinary computers in primary healthcare institutions without the need for dedicated high-end equipment.
[0116] Based on the above method, a deep learning-based system for automatically assessing adenoid hypertrophy in children using X-ray images is also disclosed. This system includes a data processing and loading module for loading X-ray images (JPG) and keypoint annotations (CSV / JSON); performing standardization processing, centering and cropping to 1200*1500 pixels, normalizing according to ImageNet parameters, and converting the format to CHW tensors; finally, data augmentation is achieved through inversion, rotation, translation, and scaling. The pre-acquired X-ray image data is divided into three parts: a training set, a validation set, and a test set; the image data in the training and validation sets are annotated; and the annotated image data in the training and validation sets is standardized.
[0117] The core architecture module of the evaluation model is constructed, and a network architecture for keypoint detection is built to achieve feature extraction, candidate box generation, and keypoint regression. An intelligent evaluation model for pediatric adenoid hypertrophy is constructed based on the ResNet-50 convolutional neural network.
[0118] The training and optimization module is responsible for the model's training loop, parameter updates, and learning rate scheduling, ensuring model convergence and optimizing detection accuracy. It trains the evaluation model using standardized training and validation sets to generate an intelligent training model for pediatric adenoid hypertrophy, and then tests the trained model using a test set, outputting the final automatic evaluation model for pediatric adenoid hypertrophy.
[0119] The prediction post-processing module filters and corrects the model's original prediction results, and outputs accurate key point coordinates.
[0120] The output module inputs real-time acquired X-ray image data into the automatic assessment model for adenoid hypertrophy in children, and automatically obtains the assessment level of adenoid hypertrophy in children.
[0121] It also includes a 3D visualization report model, which automatically generates and outputs a visualized diagnostic report containing key point predictions, grading conclusions, and image annotations.
[0122] In this example, compared with the method for automatic assessment of adenoid hypertrophy based on deep learning in MRI images (CN116630324B) in the background art, the specific differences and advantages of this solution are shown in the table below:
[0123]
[0124]
[0125] A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, invoke the steps of the method.
[0126] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a 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 described in the various embodiments of this invention. 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.
[0127] It should be understood that this solution is not limited to the specific embodiments described above. Devices and structures not described in detail herein should be understood as being implemented in a manner common to the art. Any person skilled in the art can make many possible variations and modifications to this solution, or modify it into equivalent embodiments, without departing from the scope of this solution, using the methods and techniques disclosed above. This does not affect the substantive content of this solution. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this solution, without departing from its scope, still fall within the protection scope of this solution.
Claims
1. A method for automatically assessing adenoid hypertrophy in children using X-ray images based on deep learning, characterized in that, Includes the following steps: Step S1: Divide the acquired X-ray image data into three parts: training set, validation set, and test set. Step S2: Label the image data in the training and validation sets; Step S3: Standardize the image data labeled in the training and validation sets; Step S4: Construct an intelligent assessment model for adenoid hypertrophy in children based on ResNet-50 convolutional neural network; Step S5: Use the training set and validation set obtained in the previous step to train the evaluation model and generate an intelligent training model for adenoid hypertrophy in children. Step S6: Evaluate the training model generated in step S5 using the test set obtained in step S1.
2. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 1, characterized in that, In step S2, the specific operations for data annotation are as follows: The labelme tool was used to mark five anatomical landmarks required for adenoid hypertrophy identification: the posterior border of the hard palate, the lower border of the clivus, the most prominent point of the adenoid, the junction of the pterygoid plate and the skull base, and the intersection of the surface of the lower skull base and the average posterior surface of the mandibular condyle. The image data was then converted from JPG to JSON format.
3. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 2, characterized in that, The specific operation of step S3 is as follows: Perform flip, rotate, translate, and scale operations on JSON formatted X-ray images, while simultaneously adjusting the coordinates of key points and increasing the number of samples.
4. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 3, characterized in that: Step S3 also includes a preprocessing sub-step, which performs noise reduction, smoothing, and sharpening on the image to improve image contrast, and then normalizes the pixel values to the [0-1] range through grayscale normalization.
5. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 1, characterized in that, In step S4, constructing the intelligent assessment model for adenoid hypertrophy in children includes: The feature extraction unit extracts multi-dimensional features of the image by stacking residual blocks of ResNet-50 and outputs feature maps at different levels. Anchor frame generation unit: AnchorGenerator is used to generate anchor frames with sizes ranging from 32 to 512 pixels and aspect ratios ranging from 0.25 to 4.0, for adenoid region localization. The keypoint regression unit, based on the Keypoint R-CNN architecture, predicts the keypoint locations of each target through convolution operations and non-maximum suppression.
6. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 1, characterized in that, The specific operation method of step S5 is as follows: The evaluation model was trained using the training set image data after the standardization process in step S3. The learning rate was set to 0.001, the momentum to 0.9, the weight decay to 0.0005, and the optimizer to be the SGD / Adam optimizer. The model was then validated using the validation set image data after the standardization process in step S3 using cross-validation, thus generating an intelligent training model for pediatric adenoid hypertrophy.
7. The method for automatically assessing adenoid hypertrophy in children based on X-ray films according to claim 6, characterized in that, The specific operation method of S6 is as follows: Using the test set obtained in step S1, the trained model is tested, and the A / N value is calculated using the five predicted key points to determine whether it is adenoid hypertrophy.
8. A deep learning-based automatic X-ray assessment system for adenoid hypertrophy in children, characterized by: It includes a data processing and loading module for loading X-ray images and key point annotations; it performs standardization processing, centers the images to 1200*1500 pixels, normalizes them according to ImageNet parameters, and converts them into CHW tensors; finally, it performs data augmentation through inversion, rotation, translation, and scaling. The core architecture module of the model constructs a network architecture for keypoint detection, enabling feature extraction, candidate box generation, and keypoint regression. The training and optimization module is responsible for the model's training loop, parameter updates, and learning rate scheduling, ensuring model convergence and optimizing detection accuracy. The prediction post-processing module filters and corrects the model's original prediction results and outputs accurate key point coordinates. The output module inputs real-time acquired X-ray image data into the automatic assessment model for adenoid hypertrophy in children, and automatically obtains the assessment level of adenoid hypertrophy in children.
9. The deep learning-based X-ray automatic assessment system for adenoid hypertrophy in children according to claim 8, characterized in that: It also includes a 3D visualization report model, which automatically generates and outputs a visualized diagnostic report containing key point predictions, grading conclusions, and image annotations.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions that, when executed by a processor, invoke the steps of the method according to any one of claims 1 to 7.