CN-122025105-A - CT image aided diagnosis model construction method for nasolacrimal duct obstructive disease

CN122025105ACN 122025105 ACN122025105 ACN 122025105ACN-122025105-A

Abstract

The application relates to the technical field of medical image processing and computer aided diagnosis, and discloses a nasolacrimal duct obstructive disease CT image aided diagnosis model construction method, which comprises the steps of firstly carrying out three-dimensional reconstruction and standardized pretreatment on orbital CT data; the method comprises the steps of constructing a first-level global diagnosis model, positioning focus gravity centers and outputting a cause classification result, intercepting local interesting region tensors by utilizing gravity center coordinates, constructing a second-level fine diagnosis model based on semi-supervised learning, jointly training the second-level model through supervised learning of labeled data and consistency constraint tasks of unlabeled data, outputting a smoothness classification result, and finally combining the cause and smoothness results to generate an auxiliary diagnosis report and a visual image. The application adopts a coarse-to-fine cascade architecture and a semi-supervision strategy, reduces the dependence on labeling data, and effectively improves the diagnosis precision and clinical interpretability of the nasolacrimal duct microstructure.

Inventors

WANG LIHUA
BAI FANG
ZHANG ZHISHENG

Assignees

高碑店市通远科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (10)

1. The CT image aided diagnosis model construction method for nasolacrimal duct obstructive disease is characterized by comprising the following steps: acquiring orbit CT image data, performing three-dimensional reconstruction and standardized preprocessing on the orbit CT image data to obtain preprocessed orbit CT image data, and constructing a training data set containing label data and label-free data based on the preprocessed orbit CT image data; Constructing a first-stage global diagnosis model, inputting the preprocessed orbit CT image data in the training data set into the first-stage global diagnosis model, and outputting an etiology classification result and focus barycenter coordinates; Intercepting local interesting region tensors in the preprocessed eye socket CT image data according to focus barycentric coordinates, and constructing a second-stage fine diagnosis model based on semi-supervised learning; the second level fine diagnostic model is trained jointly by supervised learning tasks on the tagged data and consistency constraint tasks on the untagged data.
2. The method for constructing a nasolacrimal duct obstructive disease CT image aided diagnosis model according to claim 1, wherein the step of three-dimensionally reconstructing and normalizing the orbital CT image data comprises: Obtaining a DICOM sequence of original orbit CT image data, and carrying out three-dimensional interpolation resampling on the DICOM sequence to generate three-dimensional volume data with isotropic voxel spacing; Acquiring preset window level parameters and window width parameters, wherein the window level parameters and the window width parameters are set based on bone window and soft tissue window characteristics; and constructing a gray mapping function based on the window level parameter and the window width parameter, mapping the voxel gray value of the three-dimensional volume data to a preset normalized numerical value interval, and generating a standardized three-dimensional input tensor serving as the preprocessed orbit CT image data.
3. The method for constructing a nasolacrimal duct obstructive disease CT image auxiliary diagnosis model according to claim 1, wherein the constructing a first-stage global diagnosis model comprises: Constructing a feature extraction backbone network for extracting global high-dimensional features of the input preprocessed orbit CT image data; Constructing parallel classification branches and regression branches at the output end of the feature extraction backbone network; the classification branches map the global high-dimensional features into primary or secondary etiology classification probability distributions; the regression branch introduces a spatial attention mechanism, carries out weighting treatment on the global high-dimensional characteristics, and regressively outputs normalized position values of the focus center in the axial direction of the nasolacrimal duct; And completing the construction of the first-stage global diagnosis model through the connection of the feature extraction backbone network, the classification branch and the regression branch.
4. A method for constructing a nasolacrimal duct obstructive disease CT image aided diagnosis model as claimed in claim 3, wherein the training process of the first-stage global diagnosis model comprises: constructing a joint loss function, wherein the joint loss function comprises a classification loss term and a regression loss term; The classification loss term is used for calculating the cross entropy error between the predicted etiology classification probability distribution and the real etiology label; the regression loss term is used for calculating a smooth L1 distance error between the predicted normalized position numerical value and the true barycentric coordinates marked by the expert; And iteratively updating parameters of the first-stage global diagnosis model based on the joint loss function.
5. The method of claim 1, wherein the step of capturing local region of interest tensors from the preprocessed orbital CT image data based on the focal barycentric coordinates comprises: acquiring the total layer number of the preprocessed orbit CT image data in the axial dimension, multiplying the focus gravity center coordinate by the total layer number and rounding to obtain a target center slice index; setting the target center slice index as a center, respectively extending a preset depth threshold upwards and downwards in the axial dimension, and determining an axial interception range; Setting the anatomical axis of the nasolacrimal duct as the center, and determining the plane interception range of the preset width and height on the cross section; And according to the axial interception range and the plane interception range, extracting a three-dimensional sub-data block from the preprocessed orbit CT image data as the local region of interest tensor.
6. The method for constructing a nasolacrimal duct obstructive disease CT image aided diagnosis model according to claim 1, wherein the constructing a second-stage fine diagnosis model based on semi-supervised learning comprises: Constructing a feature extraction network based on a hybrid architecture of a transducer and a convolutional neural network, and processing the local region of interest tensor; defining a weak enhancement operation function and a strong enhancement operation function; The weak enhancement operation function comprises random overturn and random translation transformation, and the strong enhancement operation function comprises random noise injection, contrast adjustment and affine transformation; For the unlabeled data in the training data set, respectively executing the weak enhancement operation function and the strong enhancement operation function to generate a weak enhancement view and a strong enhancement view; And executing a consistency constraint task based on the weak enhancement view and the strong enhancement view through the feature extraction network, so as to complete the construction of the second-stage fine diagnosis model based on semi-supervised learning.
7. The method for constructing a nasolacrimal duct obstructive disease CT image aided diagnosis model of claim 6, wherein the training process of the second-stage fine diagnosis model comprises: inputting the weak enhancement view into the second-stage fine diagnosis model to obtain a first prediction probability distribution; when the maximum probability value in the first prediction probability distribution exceeds a preset confidence threshold value, determining the corresponding category as a pseudo tag; Inputting the strong enhancement view into the second-stage fine diagnosis model to obtain a second prediction probability distribution; And calculating the consistency loss between the pseudo tag and the second prediction probability distribution as an optimization target of the consistency constraint task.
8. The method for constructing a nasolacrimal duct obstructive disease CT image aided diagnosis model according to claim 7, wherein the total optimized objective function of the second-stage fine diagnosis model is composed of a weighted sum of supervised and unsupervised loss terms; the supervised loss item is obtained by calculating a real hierarchical label and a model prediction result based on labeled data; The unsupervised loss term is calculated based on the consistency loss and is calculated only for unlabeled data meeting the confidence threshold; Network parameters of the second-stage fine diagnostic model are updated by minimizing the overall optimization objective function.
9. The method of claim 1, wherein the step of inputting the local region of interest tensor into the second-stage fine diagnostic model and outputting the nasolacrimal duct patency degree classification result comprises: Inputting the local region of interest tensor into the second-stage fine diagnosis model to obtain a smoothness class probability distribution; acquiring the etiology classification result output by the first-stage global diagnosis model; If the etiology classification result is judged to be secondary, generating secondary prompt information to be checked, and marking the grading result corresponding to the smoothness class probability distribution as a reference item; if the etiology classification result is judged to be primary, directly outputting a classification result corresponding to the smoothness class probability distribution as a final classification conclusion; And positioning a key slice based on the focus barycentric coordinates, and generating a key layer image overlapped with the thermodynamic diagram by using a gradient weighting type activation mapping algorithm.
10. The method of claim 9, wherein the locating key slice of focus barycentric coordinates comprises mapping the focus barycentric coordinates to voxel coordinates in a three-dimensional space, and performing slice extraction on the preprocessed orbital CT image data along three orthogonal planes of axial, coronal and sagittal positions based on the voxel coordinates, respectively, to obtain a set of two-dimensional cross-sectional images passing through the center of a focus.

Description

CT image aided diagnosis model construction method for nasolacrimal duct obstructive disease Technical Field The invention relates to the technical field of medical image processing and computer-aided diagnosis, in particular to a CT image-aided diagnosis model construction method for nasolacrimal duct obstructive disease. Background Nasolacrimal duct obstruction is a common ophthalmic disease, and if complications such as acute or chronic dacryocystitis, orbital cellulitis and the like may be caused without timely intervention, orbital CT and lacrimal passage radiography CT are main imaging means for clinically diagnosing the diseases, and doctors need to judge specific parts of obstruction, unobstructed obstruction degree, stenosis or complete obstruction, whether secondary causes such as bone destruction or space occupying lesions exist or not according to CT images, so as to formulate personalized treatment schemes. At present, the interpretation of CT images of nasolacrimal ducts mainly depends on visual observation and experience judgment of radiologists or ophthalmologists, because the nasolacrimal ducts belong to deep fine bone pipeline structures, the occupied volume proportion in three-dimensional CT images is extremely small, the surrounding is close to maxillary sinus, ethmoid sinus and complex orbital bone wall, the anatomical environment is complex, and in the conventional film reading mode, doctors need to browse dozens of or even hundreds of slices layer by layer, so that the workload is large, the time consumption is long, missed diagnosis and misdiagnosis are easy to be caused due to fatigue or experience difference, and the subjective deviation is obvious particularly when early slight stenosis is judged or primary and secondary etiology is discriminated. With the development of artificial intelligence technology, a computer-aided diagnosis system based on deep learning is gradually applied to medical image analysis, however, an automatic diagnosis research on nasolacrimal duct obstructive diseases still faces significant technical challenges, when nasolacrimal duct is used as an elongated and small-caliber anatomical structure, and when full-width three-dimensional orbit CT images are directly input into a conventional convolutional neural network for processing, in order to adapt to video memory limitation, large-scale downsampling is often required to be carried out on the images, which can lead to loss of texture details of micro-lumens, so that a model is difficult to distinguish normal physiological stenosis from pathological obstruction, and complex orbit background noise is extremely easy to interfere with feature extraction if high resolution is maintained for processing, so that focus of the model is offset. In addition, the performance of the deep learning model is highly dependent on large-scale high-quality labeling data, in the field of three-dimensional medical images, a great deal of effort and time are required for acquiring voxel-level segmentation labeling or accurate hierarchical classification labels, the data labeling cost is high, the acquisition is difficult, the existing auxiliary diagnosis method mostly adopts a fully supervised learning mode, under the condition that the labeling data are limited, the model is extremely easy to fit, the generalization capability is insufficient, and the model is difficult to adapt to clinical data under different equipment or scanning parameters, meanwhile, the existing model architecture usually only pays attention to single blocking judgment, lacks the capability of primary or secondary synchronous analysis of etiology attributes, and is difficult to meet the actual requirements of clinically on refined and structured diagnosis information. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a method for constructing a CT image auxiliary diagnosis model of nasolacrimal duct obstructive disease, which solves the problems of difficult recognition of a nasolacrimal duct microstructure, excessive dependence of diagnosis on doctor experience and high dependence of a deep learning model on large-scale labeling data in the prior art. In order to achieve the aim, the invention is realized by the following technical scheme that the CT image aided diagnosis model construction method for the nasolacrimal duct obstructive disease comprises the following steps: acquiring orbit CT image data, performing three-dimensional reconstruction and standardized preprocessing on the orbit CT image data to obtain preprocessed orbit CT image data, and constructing a training data set containing label data and label-free data based on the preprocessed orbit CT image data; Constructing a first-stage global diagnosis model, inputting the preprocessed orbit CT image data in the training data set into the first-stage global diagnosis model, and outputting an etiology classification result and focus barycenter coordinates;