Knee osteoarthritis damage prediction method based on ct images

Abstract

The application belongs to the field of medical image pattern recognition, and provides a knee osteoarthritis injury prediction method based on CT images, which comprises the following steps: step one, acquiring a group of healthy controls and a group of CT images of subjects with different degrees of knee joint damage; step two, pre-processing the X-ray images, cropping out the knee joint region, and converting the gray-scale images to L, a and b channels by using the LAB color model; step three, performing image enhancement on the images in the L channel by using adaptive histogram equalization; step four, integrating the three channels of L, a and b after enhancement; step five, performing denoising on the enhanced images; step six, marking the images of the healthy controls as 0, and marking the images of the patients with knee osteoarthritis as 1, 2, 3 and 4 according to the severity; and step seven, performing feature extraction by using a discriminative autoencoder; the method can reduce artificial errors, is an effective supplement for clinicians to judge knee osteoarthritis, and is expected to realize accurate prediction of knee osteoarthritis in areas where professional medical personnel are lacking.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image pattern recognition, in particular to a knee osteoarthritis damage prediction method based on CT images. BACKGROUND

[0002] Knee osteoarthritis, as a common chronic joint disease, has affected the quality of life of millions of people worldwide. According to epidemiological data, the incidence of knee osteoarthritis is particularly significant in the elderly population, with an estimated 5% to 10% of the global elderly population affected. In China, with the acceleration of population aging, the prevalence of knee osteoarthritis is also rising, bringing a huge economic burden to the social medical system and families. Patients often face persistent pain, joint stiffness and dysfunction, which can limit their daily activities such as walking, climbing stairs and even simple self-care. In addition, the quality of life of patients has significantly decreased, often requiring family members to provide daily care, which increases the economic burden and psychological pressure on families.

[0003] Current methods for assessing knee osteoarthritis include clinical symptom assessment, imaging examination and laboratory testing. Clinical assessment usually relies on patient complaints and physical examination results, such as knee pain, swelling and limited mobility. In terms of imaging examination, X-ray is the most commonly used tool. X-ray can clearly show degenerative changes in bone joints, such as spurs and joint space narrowing, which can help assess cartilage damage and other soft tissue problems within the joint. Currently, clinical diagnosis and treatment expect to track patients with knee osteoarthritis throughout the cycle, assess the severity of damage, develop more effective treatment plans and intervene in a timely manner to delay disease progression. However, due to the complexity of knee joint damage and the diversity of symptoms, existing diagnostic methods such as X-ray and magnetic resonance imaging mainly focus on changes in bone structure, although they can reveal structural damage to the joint, they may not be able to fully assess the actual damage to joint function.

[0004] Therefore, the present application proposes a machine learning algorithm based on CT images, which can accurately predict the grade of knee joint damage. This intelligent and highly accurate prediction method helps to predict the precise prognosis of patients and develop personalized rehabilitation programs, which can monitor disease changes over the long term, thereby improving treatment outcomes and improving the quality of life of patients. SUMMARY

[0005] To solve the above problems in the prior art, the present application proposes a knee osteoarthritis damage prediction method based on CT images. To achieve the above purpose, the present application adopts the following technical solutions:

[0006] Step one, obtaining a group of healthy controls and a group of subjects with different degrees of knee joint damage from CT images;

[0007] Step two, pre-process the X-ray image, crop the knee joint area and convert the gray image to L, a, b three channels using LAB color model;

[0008] Step three, image enhancement is performed on the L channel image using adaptive histogram equalization; Step four, integrate the enhanced L, a, b three channels;

[0009] Step five, denoising is performed on the enhanced image. Here, the fast non-local mean denoising algorithm is used, which is as follows:

[0010] Suppose the original image containing noise is v, and the denoised image is The gray value of the middle pixel point is obtained in the following way,

[0011]

[0012] Where w(x, y) represents the similarity between pixel points x and y, which reflects the distance between the rectangular neighborhood centered at x and y,

[0013]

[0014] Where Z(x) represents the normalization coefficient, and h represents the Gaussian smoothing parameter.

[0015] Step six, the image of the healthy control is marked as 0, and the image of the knee osteoarthritis patient is marked as 1, 2, 3, 4 according to the severity (1-4 represents the severity, 1 is the lightest damage, and 4 is the heaviest damage);

[0016] Step seven, feature extraction is performed using a discriminative autoencoder; the input image sample is 224x224, and the columns are concatenated into a column vector. Here, the sample set is {x i}, i = 0, 1, 2,..., 4, and the output label is {y i}, i = 0, 1, 2,..., 4. The loss function of the discriminative autoencoder model consists of three parts,

[0017] L = L AE + L distance + L weight

[0018] Where,

[0019]

[0020] And the encoder function in the autoencoder model is f enc , the decoder function is f dec , and the hidden variable is z i, N is the number of samples, I is the number of categories, theta is the number of model hyperparameters in the autoencoder, is the weight of the encoder hidden layer.

[0021]

[0022] L distance represents the distance between different categories expected by the discriminative encoder | mu i - mu j | maximum, mu i is the mean of the learned latent features in the low-dimensional space by the autoencoder model, is the variance, m (i) is the number of training samples within a category.

[0023]

[0024] L weight represents the L2 regularization parameter, K is the number of hidden layers in the autoencoder.

[0025] Step eight, output classification effect. When the discriminative autoencoder model is learned to correspond to the parameters of each category, a logistic regression model is connected to the back of the autoencoder model in each category to determine whether the input hidden variable is of the current type. The specific model is as follows:

[0026]

[0027] where f(z i ) is a linear transformation of the hidden variable, and gamma(z i ) is the output value of the Sigmod function. In the discriminative autoencoder model, the output {gamma 1 (z1), gamma 2 (z2),..., gamma N (z N )} is obtained. The category to which gamma(z i ) is the largest is taken as the final output, and is converted into labels 0, 1, 2, 3, and 4. The classification accuracy, recall rate, precision rate, and F1 score are used to measure the classification effect of the present application; the calculation method is as follows:

[0028]

[0029] BRIEF DESCRIPTION OF DRAWINGS

[0030] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0031] Figure 1The flow chart of the CT-based knee injury grade prediction method of the present application;

[0032] Figure 2 The knee image injury grade schematic diagram used in the present application;

[0033] Figure 3 The prediction model schematic diagram proposed in the present application;

[0034] Figure 4 The receiver operating characteristic curve of the model proposed in the present application for predicting 5 grades of knee. DETAILED DESCRIPTION

[0035] In order to make the purpose and technical implementation of the present application clearer, the embodiments of the present application are described in detail below, and the examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] In the description of the present application, unless otherwise specified and limited, it is necessary to explain that the terms "injury", "image", "encoder", "decoder", "connection", "model", "optimization", "training set", "test set" and the like should be understood broadly, and the specific meanings of the above terms can be understood according to specific circumstances by those skilled in the art.

[0037] As Figure 1 , the present application provides a CT-based knee injury grade prediction method, which uses CT data of 500 subjects for testing, including 145 healthy controls, 130 patients with injury grade I, 75 patients with injury grade 2, 70 patients with injury grade 3, and 80 patients with injury grade 4. The test includes the following steps:

[0038] Step one, obtain a group of healthy controls (Healthy control, HC) and a group of CT images of subjects with different degrees of knee injury, a total of 500 cases;

[0039] Step two, pre-process the X-ray image, crop the knee joint region, and convert the gray-scale image to L, a, b three channels using the LAB color model;

[0040] Step three, the image in the L channel is enhanced by adaptive histogram equalization;

[0041] Step four, integrate the enhanced L, a, b three channels;

[0042] Step five, the enhanced image is denoised. Here the denoising uses a fast non-local mean denoising algorithm, as follows:

[0043] Let v be the original image containing noise, and let v The gray value of the middle pixel point is obtained in the following manner:

[0044]

[0045] where w(x, y) represents the degree of similarity between pixel points x and y, which reflects the distance between the rectangular neighborhood centered at x and y.

[0046]

[0047] where Z(x) represents the normalization coefficient, and h represents the Gaussian smoothing parameter.

[0048] Step six, as Figure 2 , the image of the healthy control is labeled as 0, and the image of the knee osteoarthritis patient is labeled as 1, 2, 3, and 4 according to the severity (1-4 represents the severity, 1 being the least damaged and 4 being the most damaged);

[0049] Step seven, feature extraction is performed using a discriminative autoencoder; as Figure 3 , the input image sample is 224x224, and the columns are concatenated into a column vector. Here the sample set is {x i}, i = 0, 1, 2,..., 4, and the output label is {y i}, i = 0, 1, 2,..., 4. The loss function of the discriminative autoencoder model consists of three parts,

[0050] L = L AE + L distance + L weight

[0051] where,

[0052]

[0053] and the encoder function in the autoencoder model is f enc , the decoder function is f dec , the hidden variable is z i , N is the number of samples, I is the number of categories, θ is the model hyperparameter number in the autoencoder, is the weight of the encoder hidden layer.

[0054]

[0055] Ldistance represents the distance between different categories of discriminative encoder i -μ j |maximum, μ i is the mean of the learned latent features in the low-dimensional space for the auto-encoding model, is the variance, m (i) is the number of training samples within the category.

[0056]

[0057] L weight represents the L2 regularization parameter, K is the number of hidden layers in the auto-encoding

[0058] Step eight, output the classification effect. As Figure 4 , after learning the parameters corresponding to each category using the above discriminative auto-encoding model, a logistic regression model is connected to the auto-encoding model of each category to determine whether the input latent variable is of the current type. The specific model is as follows:

[0059]

[0060] where f(z i ) is a linear transformation of the latent variable, and γ(z i ) is the output value of the Sigmod function. In the discriminative auto-encoding model, the outputs {γ 1 (z1), γ 2 (z2),..., γ N (z N )} are obtained. The category with the maximum γ(z i ) is taken as the final output and converted into labels 0, 1, 2, 3, and 4. The invention uses classification accuracy, recall rate, precision rate, and F1 score to measure the classification effect; the calculation method is as follows:

[0061]

[0062] The test results are shown in the following table:

[0063]

[0064]

[0065] While the application has been described with reference to particular embodiments thereof, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present application. It will be apparent to those skilled in the art that numerous modifications can be made within the scope of the present application as defined by the appended claims. It is intended that all such modification fall within the spirit and scope of the present application. It will be understood that the features described in connection with one embodiment can be used in connection with another embodiment.

Claims

1. A knee osteoarthritis damage prediction method based on CT images, characterized by, Comprising the following steps: Step one, obtaining a group of healthy controls and a group of subjects with different degrees of knee joint damage CT images; Step two, pre-processing the CT image, cropping out the knee joint region and converting the gray image to L, a, b three channels using LAB color model; Step three, the image of L channel is enhanced by adaptive histogram equalization; Step four, integrate the enhanced L, a, b three channels; Step five, denoising the enhanced image; The denoising algorithm is as follows: Assume that the original image containing noise is , the denoised image is , The gray value of the pixel point in the middle is obtained in the following manner, ; wherein, represents the degree of similarity between the pixel points x and y, which reflects the distance between the rectangular neighborhoods centered at x and y, ; wherein denotes a normalization coefficient, h denotes a Gaussian smoothing parameter; Step six, the images of healthy controls are marked as 0, and the images of patients with knee osteoarthritis are marked as 1, 2, 3, 4 according to the severity; Step seven, feature extraction is performed using discriminative autoencoder; after learning the parameters corresponding to each category using discriminative autoencoder model, a logistic regression model is connected to the back of each category of autoencoder model to determine whether the input hidden variable is of the current type; The input image sample is 224x224, and the columns are spliced into a column vector; the sample set is , and the output label is ; the loss function of the discriminative auto-encoding model is composed of three parts, ; and the hidden variables in the autoencoder model are , N is the number of samples, I is the number of classes, is a model hyperparameter in the autoencoder. ; represents the distance between the desired classes for the discriminative encoder max, is the mean of the learned latent features in the low-dimensional space for the auto-encoding model, is the variance, is the number of training samples within a class; ; denotes the L2 regularization parameter, K is the number of hidden layers in the autoencoder; The discriminative autoencoder model is as follows: ; wherein, is a linear transformation of the latent variable, is a Sigmod function output value; in the discriminative auto-encoding model, the output ; take the class with the largest belonging as the final output, and convert it into labels 0, 1, 2, 3, 4.

2. The CT image-based knee osteoarthritis damage prediction method of claim 1, wherein, The classification accuracy, recall rate, precision rate and F1 score are used to measure the classification effect; the calculation method is as follows: 。

Images