CN-121998983-A - Adenoid hypertrophy detection method and system based on double-end knowledge distillation

CN121998983ACN 121998983 ACN121998983 ACN 121998983ACN-121998983-A

Abstract

The invention discloses a method and a system for detecting adenoid hypertrophy based on double-head knowledge distillation, and relates to the technical field of medical image processing, wherein the method comprises the steps of constructing a student segmentation network, performing joint training on the student segmentation network by utilizing a soft label generated by a teacher segmentation network and an image marked with a real label of an adenoid region, and optimizing segmentation loss between a main segmentation head output and the real label and distillation loss between a distillation head output and the soft label; inputting the oral panoramic X-ray image to be detected into a trained student segmentation network, and outputting an adenoid segmentation mask through a main segmentation head; the invention obviously improves the segmentation precision and robustness under the medical data of small samples, and realizes end-to-end classification diagnosis by fusing the segmentation mask and the clinical information.

Inventors

YANG ZHIKAI
HE JIN
LIN JINGWEN
WANG LINGFENG
YANG HAO
YANG HONGXIN
YANG TIAN

Assignees

成都信息工程大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A method for detecting adenoid hypertrophy based on double-headed knowledge distillation, comprising: constructing a student segmentation network, wherein the student segmentation network comprises an encoder and a double-head decoder connected with the encoder, the double-head decoder comprises a main segmentation head and a distillation head, and the main segmentation head and the distillation head share multi-scale characteristics output by the encoder and are independently decoded; The soft label generated by the teacher segmentation network and the image marked with the real label of the gland region are utilized to carry out joint training on the student segmentation network, so that the segmentation loss between the main segmentation head output and the real label and the distillation loss between the distillation head output and the soft label are optimized; Inputting the oral panoramic X-ray image to be detected into a trained student segmentation network, and outputting an adenoid segmentation mask through a main segmentation head; And (3) performing feature fusion on the adenoid segmentation mask and clinical information, and inputting the obtained product into a classification network to obtain an adenoid hypertrophy detection result.
2. The method for detecting adenoid hypertrophy based on double knowledge distillation as claimed in claim 1 wherein said encoder is a transducer-based encoder for extracting multi-scale features of an input image by hierarchical downsampling and outputting feature maps of different levels.
3. The method for detecting adenoid hypertrophy based on double knowledge distillation as claimed in claim 2 wherein said transducer based encoder is a Swin transducer based encoder, step down sampling is achieved by PATCH MERGING.
4. The method for detecting adenoid hypertrophy based on double-headed knowledge distillation as claimed in claim 1, wherein said double-headed decoder adopts a top-down feature fusion mode to fuse each level feature outputted from the encoder with an up-sampled high-level feature step by step, and the fusion process comprises: channel alignment and noise suppression are carried out on the low-level features output by the encoder through 1X 1 convolution, and processed low-level features are obtained; and adding the processed low-level features and the up-sampled high-level features element by element to obtain high-resolution features.
5. The method for detecting adenoid hypertrophy based on dual knowledge distillation of claim 1 comprising a multi-scale channel attention module in said dual decoder for performing attention recalibration of fusion features.
6. The method for detecting adenoid hypertrophy based on double head knowledge distillation as claimed in claim 5 wherein said multi-scale channel attention module is an MS-CAM module for performing global and local attention modeling on high resolution features, obtaining attention weights, and multiplying the high resolution features with the attention weights channel by channel to generate final segmentation; The method comprises the steps of generating a channel-level confidence coefficient by global attention modeling through global average pooling, compressing the spatial dimension of high-resolution features, and generating a spatial sensitivity attention map of the high-resolution features through 1X 1 convolution by local attention modeling.
7. The method for detecting adenoid hypertrophy based on double knowledge distillation as claimed in claim 1 wherein the total loss function of the combined training is obtained by weighted summation of segmentation loss, distillation loss and classification loss; the segmentation loss is used for supervising the alignment of the output of the main segmentation head and expert labels; the distillation loss is used to align the distillation head output with the soft prediction of the teacher split network; The classification loss is used to supervise the adenoid hypertrophy grade prediction of the classification head output.
8. The method for detecting adenoid hypertrophy based on double head knowledge distillation as claimed in claim 1 or 7 wherein the confidence filtering is performed on the soft label output by the teacher split network before calculating the distillation loss, and the sample corresponding to the soft label is incorporated into the distillation loss calculation only when the maximum probability of the soft label is greater than a preset threshold.
9. The method for detecting adenoid hypertrophy based on double knowledge distillation as claimed in claim 7 wherein said calculation of distillation loss is applicable to all input samples and said segmentation loss is applicable only to samples with true labels.
10. A double-headed knowledge distillation-based adenoid hypertrophy detection system for performing the method of any of claims 1-9 comprising: The system comprises a network construction module, a network segmentation module and a network segmentation module, wherein the network construction module is used for constructing a student segmentation network, the student segmentation network comprises an encoder and a double-head decoder connected with the encoder, and the double-head decoder comprises a main segmentation head and a distillation head; The training module is used for carrying out joint training on the student segmentation network by utilizing the soft label generated by the teacher segmentation network and the image marked with the real label of the gland region, and optimizing the segmentation loss between the main segmentation head output and the real label and the distillation loss between the distillation head output and the soft label; the segmentation reasoning module is used for inputting the oral panoramic X-ray image to be detected into a trained student segmentation network and outputting an adenoid segmentation mask through the main segmentation head; And the classification diagnosis module is used for carrying out feature fusion on the adenoid segmentation mask and clinical information and inputting the obtained result into a classification network to obtain an adenoid hypertrophy detection result.

Description

Adenoid hypertrophy detection method and system based on double-end knowledge distillation Technical Field The application relates to the field of image processing and classification, in particular to a double-head knowledge distillation-based adenoid hypertrophy detection method and system. Background Adenoids (adenoid) are a mass of lymphoid tissue located in the posterior wall of the nasopharynx and play an important role in the child's immune defenses. Its size usually reaches a limit between 6 and 10 years of age, and then gradually degenerates, with most individuals completely degenerating before puberty. However, when the adenoid body is pathologically proliferated, i.e., the adenoid body is hypertrophic (Adenoid Hypertrophy, AH), it may cause upper respiratory symptoms such as nasal obstruction, open mouth breathing, snoring in sleep, recurrent otitis media, etc. Long-term non-intervention AH may also cause characteristic craniofacial dysplasia-adenoid appearance. Epidemiological data shows that the prevalence of AH in the general population of children is about 34%, whereas in children diagnosed with malocclusions or facial aesthetic problems this proportion is as high as 42% -70%. Currently, nasal fiber endoscopy is regarded as an AH diagnostic gold standard, but its operation is invasive, costly, and depends on specialized equipment and operational experience, and is difficult to use as a routine screening tool. In contrast, the lateral head shadow measuring plate is widely applied to orthodontic treatment as a conventional, economical and noninvasive imaging examination. By measuring the ratio of the adenoid thickness to the nasopharyngeal cavity width (i.e., adenoid-nasopharyngeal ratio, adenoid/Nasopharynx ratio, AN ratio), AN effective assessment of AH can be achieved on the cephalometric slide. However, traditional AN ratio calculation is highly dependent on manually calibrating key anatomical landmark points, is cumbersome, time-consuming, and is susceptible to subjective experience of the observer, with significant intra-observer and inter-observer variability. In recent years, studies for automatically evaluating adenoid hypertrophy using artificial intelligence technology have begun to appear. For example, liu J L et al (2021) proposed an end-to-end deep learning model based on VGG-Lite to automatically determine whether there is pathological adenoid hypertrophy directly from lateral head shadow measurement X-ray films, rao Y et al (2024) designed AdeNet to predict keypoints and calculate A/N ratios by a multi-scale local attention mechanism, he Ziling (2024) proposed a full-automatic adenoid hypertrophy assessment framework incorporating multi-planar MRI information, including ADNet keypoint detection model and ASPNet segmentation model. However, the prior art still has the following disadvantages: (1) The key point detection method is sensitive to image quality, namely, anatomical key points on images lack clear and unique boundaries, especially, the edges of adenoids often appear diffuse and cloud-like in children, the most salient points are difficult to determine, and bony marks can be blurred due to immature development or image angles; (2) The data sparseness problem is that the number of high-quality pixel-level labeling samples is very small, the deep learning model has strong dependence on large-scale high-quality labeling data, and the limited samples are very easy to cause over fitting; (3) The method lacks the interpretability that the direct classification method cannot visually display the gland region, and is difficult to meet the clinical requirement on the interpretability of the diagnosis process; (4) The segmentation precision is insufficient, the traditional CNN or single-scale transducer is difficult to sense the whole form and the local structure of the adenoid body at the same time, and boundary information is easy to lose in a low-contrast X-ray image. Disclosure of Invention Aiming at the defects in the prior art, the invention provides an adenoid hypertrophy detection method and system based on double-end knowledge distillation, which take a transducer as an encoder and a segmentation network with a double-end decoder, and utilize a soft supervision signal generated by a teacher model on non-marked or weakly marked data to remarkably improve the segmentation accuracy and robustness of the model under small sample medical data, and realize end-to-end classification diagnosis through fusion of segmentation masks and clinical information. In order to achieve the aim of the invention, the invention adopts the following technical scheme: in a first aspect, the present invention provides a method for detection of adenoid hypertrophy based on double-headed knowledge distillation comprising: constructing a student segmentation network, wherein the student segmentation network comprises an encoder and a double-head decoder connected with the encoder, the double-head dec