Enhancing knee osteoarthritis detection with AI, image denoising, and optimized classification methods and the importance of physical therapy methods

Innovations of the study and contributions to literature

• KOA is a disease that occurs due to the wear of the cartilage in the knee joint. Cartilage prevents the bones in the joint from rubbing against each other, allowing them to move comfortably. However, in the case of calcification, this cartilage thins and deteriorates over time, and the bones in the joint begin to rub against each other. This leads to pain, swelling, stiffness, and limited movement.

• Therefore, detecting knee calcification with computer-aided systems at an early stage is vital. Early detection of KOA with computer-aided systems will allow early initiation of treatment.

• In this study, eight pre-trained models were used to detect KOA. These models were used for feature extraction, and the obtained features were classified in six classifiers. As a result, the highest performance was obtained in the support vector machine (SVM) classifier in the features extracted with the DenseNet201 architecture.

• A hybrid model was also developed for the early detection of knee calcification. The DenseNet201 architecture was used as the base in the developed model. In addition, a Gaussian filter was applied to the original data set to eliminate noise in the images.

• The DenseNet201 model extracts features from the original and Gaussian-applied data sets. Using the obtained feature maps, different features of the same image are blended. The optimized feature map is then classified in the SVM classifier after NCA is performed on the merged feature map to speed up and improve the model’s performance.

• The proposed model achieved competitive results and an accuracy value of 85%.

Organization of paper

Related Studies on Measuring KOA Severity Autonomously, Theoretical Background, Application Results, Discussion, and Conclusion sections are reviewed in the rest of the article.

Related studies to measure knee osteoarthritis (KOA) severity autonomously

KOA is a serious health problem that significantly negatively affects individuals’ quality of life. Therefore, automatic detection of KOA with computer-aided systems is of great importance in terms of increasing the effectiveness of diagnosis and rehabilitation processes. Various studies are in the literature in this field, and intensive research is being carried out on the development and improvement of such systems.

Deep learning has been utilized extensively lately for various applications, including natural language processing (NLP) and computer vision (CV). CNNs have demonstrated superior representation and hierarchical learning capabilities in several applications. They are also highly effective in the analysis of medical imaging. An image classification task that is very appropriate for CNNs is the automatic identification of the severity of KOA (Litjens et al., 2017).

Zhang et al. (2020) focused on estimating knee joint localization and KL degree with Residual neural network (ResNet) and Convolutional Block Attention Module (CBAM) to detect KOA severity automatically. In this study, ResNet-18 and ResNet-34 architectures were used, and it was stated that CBAM, in particular, contributed to achieving high accuracy by detecting the regions of interest in radiographic images more accurately. The accuracy rate of the model was reported as 74.81% (Zhang et al., 2020). By removing bone masks from the bone-cartilage complex, Lee, Hong & Kim (2018) carried out a cutting-edge bone-cartilage segmentation procedure to remove cartilage. Segmentation masks in several planes were averaged, and majority voting was used to accomplish 2.5D segmentation. For the femoral and tibial cartilages, the suggested BCD-Net obtained DSC values of 98.1% and 83.8%, respectively (Lee, Hong & Kim, 2018). Pedoia et al. (2019) segmented the patellar cartilage in 1,478 participants’ TSE images using U-Net CNN. To ascertain whether cartilage lesions were present according on the musculoskeletal radiologist’s judgment, the segments were examined using a specialized classification CNN. With 80% sensitivity and specificity, the deep learning approach obtained an AUC of 0.88 on the test dataset of 222 patellar cartilages (Pedoia et al., 2019). In people with or at risk for KOA, Tiulpin et al. (2018) employed a deep Siamese CNN architecture to identify the degree of OA in posterior-anterior knee radiographs using the KL system. Experienced radiologists used the OAI database to train the CNN on 6,852 knees. The deep learning approach obtained a multiclass accuracy of 67% and a weighted kappa coefficient of 0.83 on the test dataset of 3,000 knees. Additionally, it identified radiographic knee OA with KL grade 2 or higher with an AUC of 0.93 (Tiulpin et al., 2018). In this study, CNNs are also highlighted. Antony et al. (2016) presented a deep learning-based method to identify the degree of KOA in posterior-anterior knee X-rays using the KL system in individuals who already had or were at risk for developing the condition. A total of 6,224 knees were used to train the VGG-M-128 CNN, and the OAI database’s comments from seasoned musculoskeletal radiologists were consulted. The deep learning approach outperformed a Wndchrm non-CNN-based classifier with a multiclass accuracy of 67% in assigning the same KL grade as the reference standard in the test dataset of 2,668 knees. It may be possible to increase the performance of the study with more data (Antony et al., 2016). In their study for the detection of KOA, Yildirim & Mutlu (2024) performed feature extraction using textural and CNN-based methods. They eliminated unnecessary features from the feature maps they obtained with the NCA method and classified the optimized feature map in different classifiers. The researchers achieved an accuracy value of 83.6% in the proposed model (Yildirim & Mutlu, 2024).

Teoretical background

This section of the article explains the feature extraction methods used, the Gaussian method used to remove noise from the image, the NCA method used for dimension reduction, the data set, and the proposed model.

Technological infrastructure

The publicly available digital knee X-ray dataset was used during the experiments (Kaggle, 2020). This dataset has five classes. The dataset has 514, 477, 232, 221, and 206 images in normal, doubtful, mild, moderate, and severe, respectively. The Kellgen and Lawrence methods were used to grade the knee images in the dataset, and experts in the field classified them. This article used state-of-the-art deep architectures to detect KOA in wireless capsule endoscopy images.

The AlexNet architecture was proposed by Alex Krizhevsky, who gave its name to this architecture. This 25-layer architecture won the Imagenet competition 2012 using a softmax activation function (Krizhevsky, Sutskever & Hinton, 2012). The DenseNet201 architecture was developed by Huang et al. (2017). It is structurally very similar to the ResNet architecture. The difference between this architecture and the ResNet architecture is that the activation functions are combined instead of added to the following layers (Huang et al., 2017). Tan & Le (2019) developed the EfficientNetb0 architecture. The main feature of this architecture is that the depth, width, and resolution dimensions are scaled equally using a fixed coefficient. The GoogleNet architecture was developed by Szegedy et al. (2015). This architecture was designed by using parallel-connected modules instead of adding many layers on top of each other to reduce the memory usage of the network and the possibility of over-learning the network (Szegedy et al., 2015). The InceptionV3 architecture was developed by Szegedy et al. (2015) just like the GoogleNet architecture. The main feature of this model is using inception modules to increase computational efficiency without decreasing the classification accuracy (Szegedy et al., 2016). MobileNetV2 architecture was developed by Howard et al. (2017). This CNN architecture is optimized primarily for portable devices with low computational capacity (Howard et al., 2017). ResNet50 architecture was developed by He et al. (2016). This architecture uses residual connections to eliminate vanishing gradients and degrading problems. In this way, it presents a more efficient network structure (He et al., 2016). ShuffleNet architecture is also a CNN architecture developed for devices with low computational capacity, like MobileNetV2 architecture. This architecture can also be used effectively in real-time image processing applications (Zhang et al., 2018).

Generally, CNN architectures can achieve high-accuracy classification on 2D images. However, the architectures used in this study were taken as a basis, aiming to increase the accuracy values by trying to improve them. For this purpose, the Gaussian method was used to reduce the noise in the dataset’s WCE images. The Gaussian method is a method that generally tries to eliminate the noise in the image by replacing each pixel value in the image with the weighted average of the neighboring pixels (Ito & Xiong, 2000). The study used different classification algorithms accepted in the literature to classify the feature maps obtained. These algorithms are SVM (Chapelle, Haffner & Vapnik, 1999), k-nearest neighbors (KNN) (Weinberger, Blitzer & Saul, 2005), neural network (NN) (Wang & Wang, 2003), naive Bayes (NB) (Rish, 2001), logistic regression (LR) (Kleinbaum et al., 2002), decision tree (DT) (Quinlan, 1996).

In the study, feature selection was performed with the neighborhood component analysis (NCA) so that the models can work faster and produce more effective results. NCA uses a truth function during learning. This truth function rearranges the distances between classes and aims to make the examples from the same class closer and the examples from different classes farther. Thus, a feature space is obtained in which the data points are better grouped and more distinct classes are created (Goldberger et al., 2004).

Proposed model

In the proposed study for the detection of knee arthritis, WCE images were used. DenseNet201 was used as the base in the proposed model because it achieved the highest performance among the eight architectures used in the article. While the size of the train feature map obtained in the original dataset with the DenseNet201 architecture was 1,321 × 1,000, the size of the test matrix was 329 × 1,000. A Gaussian filter was applied to the original images in the proposed model to obtain high-performance metrics. In this way, it aimed to eliminate the image noise and obtain more successful features. Then, the feature maps of the improved images obtained by applying Gaussian were extracted with the DenseNet201 architecture. As in the original dataset, the size of the train feature map was 1,321 × 1,000, while the size of the test matrix was 329 × 1,000. Then, the feature maps obtained from the original and Gaussian-applied datasets were concatenated. At this stage, the aim was to bring together different features of the same image. At this stage, the size of the train feature map was 1,321 × 2,000, while the size of the test matrix was 329 × 2,000. The NCA method was used to make the developed system work faster and obtain more successful performance measurement metrics. At this stage, the size of the train feature map was 1,321 × 600, and the size of the test matrix was 329 × 600. The block diagram of the proposed system is presented in Fig. 1.

The proposed system’s performance was also compared with the results obtained by classifying the features using eight different pre-trained models in six different classifiers.

Application results

The results for the early detection of KOA with computer-aided systems were obtained in the MATLAB environment on a computer with an i7 processor, 32 GB memory, a 6 GB graphics card, and Windows 11 operating system. The models’ performance was evaluated using measures such as F1-score, Matthews correlation coefficient, accuracy, sensitivity, and specificity (Saito & Rehmsmeier, 2015).

Results obtained on pre-trained Models

In order to compare the performance of the proposed model for the detection of KOA with computer-aided systems, the results of eight different pre-trained models are presented. The most successful architecture among these models was DenseNet201, while the least successful architecture was EfficientNetb0. Confusion matrices represent classes 1-Normal, 2-Doubtful, 3-Mild, 4-Moderate, 5-Severe. Pre-trained models were run for 15 epochs and 1,230 iterations. All models were obtained using the same training parameters to compare the results successfully. Also, when the features obtained from the models were given to the classifiers, the parameters of the classifiers were kept constant. Table 1 shows the confusion matrices that were acquired in pre-trained models.

Examining Table 1, it is evident that DenseNet201 is the most effective model, with an accuracy rate of 66.26%. The DenseNet201 architecture correctly classified 218 of 329 test images while incorrectly classifying 111. The EfficientNetb0 architecture was the least successful in detecting knee arthritis. The EfficientNetb0 architecture correctly classified 128 of 329 test images while incorrectly classifying 201. The accuracy value of this model was 38.91%.

Classification of feature maps obtained in DenseNet201 architecture using classifiers

Table 2 shows the accuracy values derived from classifying the feature maps from eight separate models in six different classifiers to compare the performance of the suggested model.

The DenseNet201 architecture was the foundation for the suggested model since it produced the best performance rate among the pre-trained models. The features acquired in eight distinct models were optimized using the NCA approach, and they were subsequently categorized into six different classifiers to assess the effectiveness of the suggested model. One thousand features were acquired for every image. Then, 600 features were selected for each image from these features with the NCA method. In the experiments conducted at this stage, the cross-validation value was selected as five. The obtained accuracy values are presented in Table 3.

When Table 2 is examined, the Cubic SVM classifier achieved the highest accuracy value, 79.6%, after the features obtained with the DenseNet201 architecture were optimized with NCA. Table 4 gives the confusion matrices obtained by classifying the features obtained in the DenseNet201 architecture into six different classifiers.

Cubic SVM correctly classified 1,313 of the 1,650 test images, while 337 were incorrectly classified. Similarly, the least successful classifier was Fine Tree, with 55.1% accuracy. Fine Tree correctly classified 909 of the 1,650 test images and incorrectly classified 741.

Results of the proposed model

In the proposed model, for the detection of knee arthritis, feature maps obtained from both the original and Gaussian applied datasets with the DenseNet201 architecture were combined to bring together different features of the same image. After this merging process, 2,000 features were obtained for each image. Then, 600 features were selected using the NCA method to increase the performance of the proposed model. Six distinct classifiers were used to classify the optimized feature map. Table 5 shows the accuracy values of the suggested model derived from six distinct classifiers.

With an 85% success rate, the SVM classifier is the most effective classifier, according to Table 5. Neural network follows the SVM classifier at 79.9%, KNN at 78.5%, Logistic Regression at 75.6%, Naive Bayes at 64.1%, and Fine Tree at 53.8%. Since the highest performance values in the proposed model are reached in the SVM classifier, SVM is preferred as the classifier in the proposed model. The confusion matrix of the proposed model is in Table 6.

Table 6 shows that the suggested model correctly identified 1,402 of the 1,650 test images, while 248 were incorrectly identified. In the Normal class, 461 of 514 test images were correctly classified, and 53 were incorrectly classified. In the Doubtful class, 401 of 477 test images were correctly classified, and 76 were incorrectly classified. In the Mild class, 162 of 232 test images were correctly classified, and 70 were incorrectly classified. In the Moderate class, 197 of 221 test images were correctly classified, and 24 were incorrectly classified. Finally, 181 of 206 test images were correctly classified in the Severe class, and 25 were incorrectly classified. Table 7 shows the suggested model’s performance measuring metrics.