A method and system for detecting adenoid hypertrophy based on dual-head knowledge distillation
By using a Transformer-based dual-head knowledge distillation network and leveraging soft-supervised signals and multi-scale channel attention modules, the problems of key point detection and data sparsity in adenoid hypertrophy detection are solved, achieving high-precision and interpretable adenoid hypertrophy detection suitable for non-invasive diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU UNIV OF INFORMATION TECH
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies for detecting adenoid hypertrophy suffer from problems such as sensitivity to image quality in key point detection, data sparsity, lack of interpretability, and insufficient segmentation accuracy. In particular, the edges of adenoids are blurred in pediatric images, making it difficult to meet the requirements for non-invasive and interpretable diagnosis.
A segmentation network based on Transformer encoder and dual-head decoder is adopted, and joint training is performed using soft supervision signals generated by the teacher model. Through multi-scale channel attention module and knowledge distillation mechanism, the segmentation accuracy and robustness of the model under small sample data are improved, and end-to-end classification diagnosis is performed by combining clinical information.
It significantly improves the segmentation accuracy and model robustness of adenoid hypertrophy detection, provides interpretable adenoid region display, meets clinical diagnostic needs, and enhances the ability to perceive boundary information in low-contrast X-ray images.
Smart Images
Figure CN121998983B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of image processing and classification, and in particular to a method and system for detecting adenoid hypertrophy based on dual-head knowledge distillation. Background Technology
[0002] The adenoids are a mass of lymphoid tissue located on the posterior wall of the nasopharynx, playing a vital role in children's immune defense. Their size typically reaches its limit between 6 and 10 years of age, after which they gradually regress, with most individuals completely regressing before puberty. However, when the adenoids undergo pathological hyperplasia, known as adenoid hypertrophy (AH), it can lead to upper respiratory symptoms such as nasal congestion, mouth breathing, snoring during sleep, and recurrent otitis media. Long-term untreated AH can also cause a characteristic craniofacial developmental abnormality—"adenoid facies." Epidemiological data show that the prevalence of AH in the general child population is approximately 34%, while this proportion is as high as 42%–70% in children seeking medical attention for malocclusion or facial aesthetic problems.
[0003] Currently, nasal fiberoptic endoscopy is considered the gold standard for diagnosing adenoid hemorrhage (AH). However, its invasiveness, high cost, and reliance on specialized equipment and operator experience make it unsuitable as a routine screening method. In contrast, lateral cephalometric radiography, as a routine, economical, and non-invasive imaging examination, is widely used in orthodontic treatment. By measuring the ratio of adenoid thickness to nasopharyngeal width (i.e., the adenoid / nasopharyngeal ratio, AN ratio), AH can be effectively assessed on cephalometric radiographs. However, traditional AN ratio calculation heavily relies on manual marking of key anatomical landmarks, a cumbersome and time-consuming process that is easily influenced by the observer's subjective experience, exhibiting significant intra- and inter-observer variability.
[0004] In recent years, research has begun to emerge on the use of artificial intelligence technology for the automatic assessment of adenoid hypertrophy. For example, Liu JL et al. (2021) proposed an end-to-end deep learning model based on VGG-Lite to automatically determine the presence of pathological adenoid hypertrophy directly from lateral cephalometric X-ray images; Rao Y et al. (2024) designed AdeNet, which predicts key points and calculates the A / N ratio through a multi-scale local attention mechanism; and He Ziling (2024) proposed a fully automatic adenoid hypertrophy assessment framework that integrates multiplanar MRI information, including the ADNet key point detection model and the ASPNet segmentation model.
[0005] However, existing technologies still have the following shortcomings: (1) Key point detection methods are sensitive to image quality: Anatomical key points lack clear and unique boundaries on images, especially in children, the edge of the adenoids is often diffuse and hazy, making it difficult to determine the most prominent point, and bony landmarks may be blurred due to immature development or image angle. (2) Data sparsity problem: There are very few high-quality pixel-level labeled samples, while deep learning models are highly dependent on large-scale high-quality labeled data. Limited samples are very likely to cause overfitting. (3) Lack of interpretability: Direct classification methods cannot visually display the adenoid region, making it difficult to meet the clinical requirements for interpretability of the diagnostic process; (4) Insufficient segmentation accuracy: Traditional CNN or single-scale Transformer cannot simultaneously perceive the overall morphology and local structure of adenoids, and are prone to losing boundary information in low-contrast X-ray images. Summary of the Invention
[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a method and system for detecting adenoid hypertrophy based on dual-head knowledge distillation. The system uses a segmentation network with a Transformer encoder and a dual-head decoder. It utilizes soft supervision signals generated by a teacher model on unlabeled or weakly labeled data to significantly improve the segmentation accuracy and robustness of the model on small-sample medical data. Furthermore, it achieves end-to-end classification diagnosis by fusing segmentation masks with clinical information.
[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for detecting adenoid hypertrophy based on dual-head knowledge distillation, comprising: A student segmentation network is constructed, which includes an encoder and a dual-head decoder connected to the encoder. The dual-head decoder includes a main segmentation head and a distillation head, which share the multi-scale features output by the encoder and decode them independently. The student segmentation network was jointly trained using soft labels generated by the teacher segmentation network and images labeled with real labels of the adenoid region, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The panoramic X-ray image of the oral cavity to be detected is input into the trained student segmentation network, and the adenoid segmentation mask is output through the main segmentation head. After fusing the adenoid segmentation mask with clinical information, the data is input into a classification network to obtain the adenoid hypertrophy detection results.
[0008] Preferably, the encoder is a Transformer-based encoder used to extract multi-scale features of the input image through hierarchical downsampling and output feature maps at different levels.
[0009] Preferably, the Transformer-based encoder is a Swing Transformer-based encoder that uses Patch Merging to achieve progressive downsampling.
[0010] Preferably, the dual-head decoder employs a top-down feature fusion method, fusing the features output from each level of the encoder with the upsampled higher-level features level by level. The fusion process includes: The low-level features output by the encoder are channel aligned and noise suppressed by 1×1 convolution to obtain the processed low-level features. The processed low-level features are added element-wise to the upsampled high-level features to obtain high-resolution features.
[0011] Preferably, the dual-head decoder includes a multi-scale channel attention module for attention recalibration of the fused features.
[0012] Preferably, the multi-scale channel attention module is an MS-CAM module, which is used to perform global and local attention modeling on high-resolution features, obtain attention weights, and multiply the high-resolution features and attention weights channel by channel to generate the final segmentation; Among them, global attention modeling compresses the spatial dimension of high-resolution features through global average pooling and generates channel-level confidence after 1×1 convolution; local attention modeling generates spatially sensitive attention maps of high-resolution features through 1×1 convolution.
[0013] Preferably, the total loss function for joint training is obtained by weighted summation of segmentation loss, distillation loss, and classification loss; The segmentation loss is used to monitor the alignment of the main segmentation head's output with expert annotations; The distillation loss is used to align the distillation head output with the soft prediction of the teacher segmentation network; The classification loss is used to supervise the prediction of adenoid hypertrophy level output by the classification head.
[0014] Preferably, before calculating the distillation loss, the soft labels output by the teacher segmentation network are filtered by confidence. Only when the maximum probability of a soft label is greater than a preset threshold is the sample corresponding to that soft label included in the distillation loss calculation.
[0015] In a second aspect, the present invention also provides an adenoid hypertrophy detection system based on dual-head knowledge distillation, for performing the method described in the first aspect, comprising: A network construction module is used to construct a student segmentation network, the student segmentation network including an encoder and a dual-head decoder connected to the encoder, the dual-head decoder including a main segmentation head and a distillation head; The training module is used to jointly train the student segmentation network using soft labels generated by the teacher segmentation network and images labeled with real labels of adenoid regions, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The segmentation inference module is used to input the oral panoramic X-ray image to be detected into the trained student segmentation network, and output the adenoid segmentation mask through the main segmentation head; The classification and diagnosis module is used to fuse the adenoid segmentation mask with clinical information and then input it into the classification network to obtain the adenoid hypertrophy detection results.
[0016] The beneficial effects of this invention are as follows: 1. This invention utilizes a dual-head knowledge distillation mechanism to provide effective supervision on unlabeled samples using soft labels generated by a teacher model, significantly alleviating the problem of sparsity in medical image annotation data and improving the segmentation accuracy and robustness of the model under small sample conditions; 2. This invention uses a multi-scale channel attention module to perform global and local attention recalibration on fused features, effectively enhancing the response to blurred adenoid boundaries and improving segmentation accuracy; 3. This invention visualizes the adenoid region through pixel-level segmentation masking and combines it with clinical information to perform multimodal feature fusion, outputting the hypertrophy level, which not only meets the needs of clinical diagnosis but also provides interpretability. Attached Figure Description
[0017] Figure 1 This invention provides a method for detecting adenoid hypertrophy based on dual-head knowledge distillation in some embodiments. Figure 2 This is a schematic diagram of the TANet classification module provided in some embodiments of the present invention; Figure 3 This is a schematic diagram of the Swin-DCAM module provided in some embodiments of the present invention; Figure 4 This is a schematic diagram of a classification module provided in some embodiments of the present invention; Figure 5 This is a schematic diagram of the LWA module provided in some embodiments of the present invention. Detailed Implementation
[0018] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0019] In the description of this invention, the terms "comprising," "including," and any variations thereof are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes 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 process, method, product, or apparatus.
[0020] like Figure 1 As shown, this embodiment provides a method for detecting adenoid hypertrophy based on dual-head knowledge distillation, including: A student segmentation network is constructed, the student segmentation network including an encoder and a dual-head decoder connected to the encoder, the dual-head decoder including a main segmentation head and a distillation head; The student segmentation network was jointly trained using soft labels generated by the teacher segmentation network and images labeled with real labels of the adenoid region, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The panoramic X-ray image of the oral cavity to be detected is input into the trained student segmentation network, and the adenoid segmentation mask is output through the main segmentation head. After fusing the adenoid segmentation mask with clinical information, the data is input into a classification network to obtain the adenoid hypertrophy detection results.
[0021] As one possible implementation, the encoder is a Transformer-based encoder used to extract multi-scale features of the input image through hierarchical downsampling and output feature maps at different levels.
[0022] As a preferred implementation, the classification network adopts the TANet classification module, the structure of which is shown in the figure below. Figure 2 As shown.
[0023] As a preferred implementation, the Transformer-based encoder is a SwingTransformer-based encoder, which implements progressive downsampling through Patch Merging. To this end, the present invention introduces, for example... Figure 3 The Swin-based Dual-head Context-Aware Module (Swin-based Dual-head Context-Aware Module) shown is used to achieve efficient and robust adenoid region segmentation and disease state discrimination, and its generalization ability is particularly optimized in data-sparse scenarios.
[0024] The dual-head decoder employs a top-down feature fusion approach, fusing the features output from each level of the encoder with the upsampled higher-level features step by step. The fusion process includes: The low-level features output by the encoder are channel aligned and noise suppressed by 1×1 convolution to obtain the processed low-level features. The processed low-level features are added element-wise to the upsampled high-level features to obtain high-resolution features.
[0025] Specifically, when the Swin-DCAM module operates in the network, it first receives multi-scale feature maps output from the Swin Transformer encoder. ,in, This is the highest semantic level feature; Will Input the pyramid pooling module (PPM) to obtain multi-scale contextual information and obtain enhanced features. ; At the same time, They are used as jump connection signals and fused with the corresponding hierarchical features in the top-down upsampling path.
[0026] Traditional CNNs or single-scale Transformers extract features only at a fixed resolution, making it difficult to simultaneously perceive the overall morphology (high-level semantics) and local structure (low-level details) of the adenoids. Therefore, this invention introduces hierarchical encoding using Swin Transformer + Patch Merging. By utilizing the four-stage hierarchical structure of Swin Transformer and implementing progressive downsampling through Patch Merging, the model simultaneously possesses both "macroscopic judgment" and "microscopic perception".
[0027] Even with shallow features (F1), they still contain a lot of background noise; and traditional attention methods (such as SE) dilute the response of small targets due to the use of GAP, causing blurred boundary regions to be ignored. To address this, this invention introduces a multi-scale channel attention module (MS-CAM) to obtain channel-level confidence while preserving spatial dimensions, generating pixel-level attention, recalibrating the input features, and enhancing the response of low-level details.
[0028] Simply splicing or adding high-level and low-level features can easily lead to a semantic gap: high-level features have strong but coarse semantics, while low-level features have many details but are noisy.
[0029] Therefore, this invention employs top-down fusion and lightweight skip connections to accurately convey details, specifically: Starting from F4, upsampling is performed stepwise (×2), and then fused with F3→F2→F1; the encoder outputs F1~F3 are first subjected to 1×1 convolutional projection (channel alignment + denoising), and then added to the upsampled features:
[0030] in, This indicates that the network at the current level has already incorporated features (usually referring to shallow layers). This indicates that the next layer of the network has incorporated features (usually referring to deeper layers). Indicates an upsampling operation. This represents a 1×1 convolution, which is then further refined using the MS-CAM module at a high-resolution layer (H / 4), preserving the original details of F1 while injecting semantic guidance from F4.
[0031] Subsequently, in the main path of the decoder, the fused high-resolution features It is sent to the MS-CAM module, and the MS-CAM module... Simultaneous global and local attention modeling: The global branch compresses the spatial dimension using global average pooling (GAP) and generates channel-level confidence after a 1×1 convolution; the local branch directly generates a spatially sensitive attention map using a 1×1 convolution; the two are element-wise summed and activated by a sigmoid function to obtain the comprehensive attention weights. Subsequently, With weight Channel-by-channel multiplication yields the recalibration characteristics. After 4x bilinear upsampling and 1×1 convolution, the final segmentation logits are output, as shown in the following formula:
[0032]
[0033]
[0034] in, It is the Sigmoid activation function. This represents element-wise multiplication. This indicates a global average pooling operation. This indicates a 4x upsampling.
[0035] Since models struggle to learn reliable boundary patterns from a small number of samples when labeled data is sparse, this invention also introduces a two-head knowledge distillation: the teacher's prior knowledge is used to compensate for missing details. The teacher model (pre-trained U-Net++) has learned robust boundary representations on large-scale data; the student model aligns the soft probability graph (containing sub-pixel confidence) of the teacher's output through KL loss; although the distillation head is lightweight (Lightweight Attention), it can still learn the teacher's uncertainty modeling for fuzzy regions.
[0036] By using dual-head knowledge distillation, even if a case lacks expert annotation, effective supervision can be provided as long as the teacher model has high confidence, significantly improving the generalization ability for small samples.
[0037] like Figure 4 As shown, in the method proposed in this invention, the training process employs a multi-task joint optimization strategy, comprising three parts: main segmentation supervision, knowledge distillation supervision, and classification diagnosis supervision. The total loss function is composed of the following three weighted terms:
[0038] in, All are hyperparameter weights, and the loss for each part is defined as follows: Segmentation loss Used to monitor the output logits_main of Decoder A (main splitter head) and expert annotations. Alignment, for Dice loss and cross-entropy loss The weighted sum is calculated as follows:
[0039] in, The segmentation probability distribution output by the main segmentation head. , The expression for mitigating foreground-background imbalance in medical images is:
[0040] in, This is a smoothing term.
[0041] Distillation loss This is used to align the Decoder B (distillation head) output `logits_distill` with the soft predictions of the Teacher model. It is calculated using KL divergence, and the expression is as follows:
[0042] in, For teachers, segment the network output. The output of the network distillation head is segmented for students. is the distillation temperature, and || is the conditional separator for the KL divergence. Represents the output of the teacher segmentation network In Softmax, a smoothed distribution is used for the distillation temperature. This represents the output of the student segmentation network distillation head. In Softmax, a smoothed distribution is used for the distillation temperature. Let KL divergence be expressed as: Based on KL divergence, this invention designs a confidence filtering mechanism, which only filters confidence when... Only then was the sample included. Calculations are performed to avoid interference from low-quality pseudo-labels.
[0043] Classification and diagnosis of loss The adenoid hypertrophy grade prediction used to supervise the output of the TANet classification head is calculated as follows:
[0044] in, The output probability distribution of the classification module. It is a four-category clinical label (normal / mild / moderate / severe); its input features are derived from the fusion of the segmentation mask (logits_main is used to obtain a probability map through Sigmoid) and clinical information (age, symptoms, etc.).
[0045] Specifically, the calculation of the distillation loss applies to all input samples, including those without real labels; the segmentation loss applies only to samples with real labels.
[0046] The design of the aforementioned loss function enables the network to simultaneously perceive the overall morphological semantics of the adenoids and local boundary details: the global path ensures a reliable judgment of the existence of the adenoids, while the local path enhances the response to blurred edges and small protrusions. Especially under conditions of low contrast in medical X-ray images and sparse labeled data, MS-CAM effectively suppresses background noise and improves the quality of low-level features through a soft weighting mechanism, thereby significantly improving the adenoid segmentation accuracy.
[0047] Furthermore, the distillation head in the Swin-DCAM module uses a skip connection between the LWA module and a 1×1 convolutional projection. It shares encoder features with the main segmentation head but decodes independently, serving only to receive soft-label supervision from the teacher model. This enhances the model's generalization ability during training and can be completely removed during inference without affecting deployment efficiency. The LWA module is described as follows: Figure 5 As shown.
[0048] To achieve the method described in this invention, this invention also provides an adenoid hypertrophy detection system based on dual-head knowledge distillation, comprising: A network construction module is used to construct a student segmentation network, the student segmentation network including an encoder and a dual-head decoder connected to the encoder, the dual-head decoder including a main segmentation head and a distillation head; The training module is used to jointly train the student segmentation network using soft labels generated by the teacher segmentation network and images labeled with real labels of adenoid regions, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The segmentation inference module is used to input the oral panoramic X-ray image to be detected into the trained student segmentation network, and output the adenoid segmentation mask through the main segmentation head; The classification and diagnosis module is used to fuse the adenoid segmentation mask with clinical information and then input it into the classification network to obtain the adenoid hypertrophy detection results.
[0049] 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 method for detecting adenoid hypertrophy based on dual-head knowledge distillation, characterized in that, include: A student segmentation network is constructed, which includes an encoder and a dual-head decoder connected to the encoder. The dual-head decoder includes a main segmentation head and a distillation head, which share the multi-scale features output by the encoder and decode them independently. The dual-head decoder includes a multi-scale channel attention module for attention recalibration of fused features; the multi-scale channel attention module is an MS-CAM module, which is used to perform global and local attention modeling on high-resolution features, obtain attention weights, and multiply the high-resolution features and attention weights channel by channel to generate the final segmentation; Among them, global attention modeling compresses the spatial dimension of high-resolution features through global average pooling and generates channel-level confidence after 1×1 convolution; local attention modeling generates spatially sensitive attention maps of high-resolution features through 1×1 convolution. The student segmentation network was jointly trained using soft labels generated by the teacher segmentation network and images labeled with real labels of the adenoid region, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The panoramic X-ray image of the oral cavity to be detected is input into the trained student segmentation network, and the adenoid segmentation mask is output through the main segmentation head. After fusing the adenoid segmentation mask with clinical information, the results are input into a classification network to obtain the adenoid hypertrophy detection results.
2. The method for detecting adenoid hypertrophy based on dual-head knowledge distillation according to claim 1, characterized in that, The encoder is a Transformer-based encoder used to extract multi-scale features of the input image through hierarchical downsampling and output feature maps at different levels.
3. The adenoid hypertrophy detection method based on dual-head knowledge distillation according to claim 2, characterized in that, The Transformer-based encoder is a Swing Transformer-based encoder that uses Patch Merging to achieve progressive downsampling.
4. The method for detecting adenoid hypertrophy based on dual-head knowledge distillation according to claim 1, characterized in that, The dual-head decoder employs a top-down feature fusion approach, progressively fusing the features output from each level of the encoder with the upsampled higher-level features. The fusion process includes: The low-level features output by the encoder are channel aligned and noise suppressed by 1×1 convolution to obtain the processed low-level features. The processed low-level features are added element-wise to the upsampled high-level features to obtain high-resolution features.
5. The method for detecting adenoid hypertrophy based on dual-head knowledge distillation according to claim 1, characterized in that, The total loss function for joint training is obtained by weighted summation of segmentation loss, distillation loss, and classification loss; The segmentation loss is used to monitor the alignment of the main segmentation head's output with expert annotations; The distillation loss is used to align the distillation head output with the soft prediction of the teacher segmentation network; The classification loss is used to supervise the prediction of adenoid hypertrophy level output by the classification head.
6. The method for detecting adenoid hypertrophy based on dual-head knowledge distillation according to claim 1 or 5, characterized in that, Before calculating the distillation loss, the soft labels output by the teacher segmentation network are filtered by confidence. Only when the maximum probability of a soft label is greater than a preset threshold is the sample corresponding to that soft label included in the distillation loss calculation.
7. The method for detecting adenoid hypertrophy based on dual-head knowledge distillation according to claim 5, characterized in that, The calculation of the distillation loss applies to all input samples, while the segmentation loss applies only to samples with true labels.
8. A system for detecting adenoid hypertrophy based on dual-head knowledge distillation, characterized in that, For performing the method according to any one of claims 1-7, comprising: A network construction module is used to construct a student segmentation network, the student segmentation network including an encoder and a dual-head decoder connected to the encoder, the dual-head decoder including a main segmentation head and a distillation head; The training module is used to jointly train the student segmentation network using soft labels generated by the teacher segmentation network and images labeled with real labels of adenoid regions, thereby optimizing the segmentation loss between the output of the main segmentation head and the real labels, as well as the distillation loss between the output of the distillation head and the soft labels. The segmentation inference module is used to input the oral panoramic X-ray image to be detected into the trained student segmentation network, and output the adenoid segmentation mask through the main segmentation head; The classification and diagnosis module is used to fuse the adenoid segmentation mask with clinical information and then input it into the classification network to obtain the adenoid hypertrophy detection results.