Automatic Classification of the Severity of Knee Osteoarthritis Using Enhanced Image Sharpening and CNN

Abstract

Knee osteoarthritis is a significant cause of physical inactivity and disability. Early detection and treatment of osteoarthritis (OA) degeneration can decrease its course. Physicians’ scores may differ significantly amongst interpreters and are greatly influenced by personal experience based solely on visual assessment. Deep convolutional neural networks (CNN) in conjunction with the Kellgren–Lawrence (KL) grading system are used to assess the severity of OA in the knee. Recent research applied for knee osteoarthritis using machine learning and deep learning results are not encouraging. One of the major reasons for this was that the images taken are not pre-processed in the correct way. Hence, feature extraction using deep learning was not great, thus impacting the overall performance of the model. Image sharpening, a type of image filtering, was required to improve image clarity due to noise in knee X-ray images. The assessment used baseline X-ray images from the Osteoarthritis Initiative (OAI). On enhanced images acquired utilizing the image sharpening process, we achieved a mean accuracy of 91.03%, an improvement of 19.03% over the earlier accuracy of 72% by using the original knee X-ray images for the detection of OA with five gradings. The image sharpening method is used to advance knee joint recognition and knee KL grading.

1. Introduction

An estimated one in three persons will suffer from knee osteoarthritis (OA) at some point in their lives [1,2]. More than half of the people 65 and older have signs of OA, including at that one joint. More than a quarter of Americans will reach the age of 65 by 2030, putting them at risk of developing OA [3]. Knee osteoarthritis (OA) has a significant impact on the quality of life of old age people. Even worse, no treatment has emerged to date which can reduce or completely stop the advancement of knee OA’s degenerative structural changes. OA patients’ quality of life can be improved and slowed down by early detection and treatment, but there are hazards involved. OA of the knee is characterized by Joint Space Narrowing (JSN), subchondral sclerosis, and the development of osteophytes. Knee joints’ 3D architecture can be seen in MRI images.

For regular knee OA diagnosis, MRI is unsuitable since it is only available at big medical centers and is costly. X-rays have long been considered the gold standard for diagnosing osteoarthritis of the knee due to their high levels of reliability, low cost, and ubiquitous availability. According to WHO guidelines, the Kellgren and Lawrence (KL) severity grading system was adopted in 1961 [4]. It is divided into five classes, from 0 to 4, according to the KL system’s knee OA severity. For each grade level, the data are shown in Figure 1, along with the students who met the criterion for inclusion. In most cases, doctors examine a scanned X-ray image of the knees and quickly assign a KL grade to each joint. The accuracy of a doctor’s diagnosis is heavily dependent on his or her expertise and attention to detail. Furthermore, the KL grading standard is rife with ambiguity. For example, the criteria for KL grade 1 include the possibility of osteophytic lipping and a low probability of JSN. Even the same doctor may award a separate KL grade to the same knee joint at various periods. A study by Culvenor et al. [5] found a KL interrater reliability range of 0.67–0.73. Because of the ambiguous nature of the KL grade criterion, we believe that physicians’ grading of the knee joint is not as reliable as it could be [6].

The grade of a knee joint may be incorrectly changed from a lower one (for example, from grade 1 to grade 2) to a higher one (e.g., from grade 1 to grade 4). The correctness of grading should be the only criterion for evaluation. Mean absolute error (MAE), which is modelled by MAE used in age estimate assessments, is employed as a substitute metric for knee KL grade categorization evaluation. [7]. Because knee OA is so common, it is vital to precisely identify it and determine how severe it is.

Full-automation knee severity grading does not become fatigued over a long evaluation period and may offer objective, reliable projections. One must first recognize the knee joint in a raw X-ray picture and then categorize it into one of five KL categories in order to effectively predict the severity of knee osteoarthritis (OA) based just on the knee joint. In recent years, knee joint recognition and KL grade classification algorithms have been created. In order to recognize knee joints, Shamir et al. [8] employ a template matching technique and a sliding window strategy. Images of 20 knee joints of 15 × 15 pixels each are moved to the present X-ray picture, and the Euclidean distances between them are calculated. It is determined that the knee joint is located at the lowest Euclidean distance possible by looking at the window [9].

A linear support vector machine (SVM) is also employed by Antony et al. [10] to identify knee joints by using Sobel vertical image gradients as features, generated by horizontal edges in knee joint photos. By using entire radiographs as input, K. Thomas et al. developed an end-to-end easily understandable model that is as accurate as musculoskeletal radiologists and does not necessitate any human image pre-processing steps. Data relevant to the treatment of patients were used to make the model’s predictions [11]. In the first place, Tiulpin et al. [12] make suggestions for the knee joint area in their study. Due to the patella, they see an increased intensity, follow by a fast reduction in intensity because of the distance between the knee joints. As a next step, a linear SVM classifier is applied to the knee joint proposals using Histogram of Oriented Gradients (HoG) features [13]. Using a deep neural network (DNN), S. Abdulla et al. [14] describe a new approach to finding knee joints and achieving cutting-edge performance. Semantic segmentation is the primary focus of the FCN-based approach [15]. The identification of the knee joint may benefit from a more specialized CNN detection architecture.

Knee KL grade classification was proposed by Shamir et al. in early 2009, based on hand-crafted qualities such as texturing, Chebyshev statistic, and Haralick features [8]. A number of deep learning-based techniques have been attempted in the last three years by Antony et al. [10,15]. KL grade classification is seen as a regression problem by [10]. The mean squared loss is used to fine-tune BVLC CafeNet for knee KL grade classification. By combining cross-entropy and mean squared losses, they come up with a new CNN model and improve its performance. There is an algorithm called the Deep Siamese CNN that was created by Tiulpin and his colleagues [12]. Selecting sub-areas of the knee joint and then combining predictions from the specified regions are both performed using this tool.

A diverse variety of visual tasks, including image classification and object detection and segmentation, may now be handled by deep learning-based algorithms [16]. Detection and segmentation of cells [17], mitosis [18], white matter lesion segmentation [19], and retinal blood vessel segmentation [19] are just a few of the applications for deep learning in medical image analysis that have popped up in the last five years. Many researchers use algorithms based on deep learning for analyzing osteoarthritis in the knee. Nevertheless, improvements must be made to knee analyses.

The implementation of a more effective loss function can enhance knee KL grading performance. It is in this study that we employ Inception ResNet V2 CNN models sequentially to automatically assess knee OA severity and get the best results. After collecting the data, we initially applied pre-processing procedures. They applied novel image enhancement techniques using Open CV to further enhance the image quality of the knee X-ray dataset along with contour detection and segmentation methods to detect the region of interest (ROI) of knee joints. Then, the Inception ResNet V2 model of CNN architecture with parameter hyper-tuning and customized hidden layers with available features was applied to knee X-ray images for KL Grading. The knee KL grading challenge puts the most common CNN models to the test, and the results reveal that, using the Inception ResNet V2 model, the suggested methods provide the total knee severity grading performance. A knee X-ray image is shown in Figure 1 as part of a pipeline for rating severity.

A general block diagram of the proposed technique is shown below in Figure 2.

2. Related Work

When compared to other forms of OA, knee osteoarthritis (KOA) is more common. KOA is caused by a complicated interaction between mechanical forces, local inflammation, joint integrity, metabolic processes, and hereditary susceptibility. The risk factors for KOA include being overweight, becoming older, and being injured [1]. Poor quality of life, limited levels of psychology, and social tiredness as a result of low public involvement are the primary repercussions of the particular condition [2]. The disease mechanism of arthritis is poorly known because of its complex character, and prediction and diagnostic methods are currently in development.

Both the diagnosis and management of KOA illness provide difficulties for medical science. There has been a rise in the number of KOA studies using big data and AI analytics because of the exponential growth in the quantity of available data. The literature analysis revealed a number of methods wherein machine learning (ML) models were utilized to foretell the growth of KOA [3].

By utilizing five distinct indicators of incident knee OA, including medial joint space narrowing, Lazzarini et al. [4] analyzed the impact of several factors (including biomarkers) within the prediction models in overweight and obese women (JSN). This article set out to identify and evaluate new biomarkers for KOA.

Self-reported knee pain and radiographic evaluations of joint space narrowing were utilized by Halilaj et al. [5] to describe distinct OA progression clusters and develop models for early prediction of these clusters. Further, Pedoia et al. [6] used multi-dimensional MRI and biomechanics to establish a multi-dimensional platform for better OA outcome prediction and patient sub classification. This method pioneered the combination of compositional imaging with skeletal biomechanics on a wide scale. Using the Kellgren and Lawrence grading method, Abedin et al. [7] compared the accuracy of a statistical model built from patient questionnaire data to that of a model built from X-ray images for making predictions. Comparable accuracy was found between the two methods, leading researchers to propose building a model that incorporates data from both the patient’s questionnaire and the X-rays. Widera et al. [8] also explored a multi-classifier issue for predicting KOA development using clinical data and X-ray image evaluation metrics, where several algorithms and learning process configurations were examined. With the proposed strategy, the percentage of patients who make no improvement drops by 20–25%. So, further research and methods for identifying risks that lead to solid instruments for forecasting KOA are required. A brief literature survey is shown in

Table 1. Literature Survey.

Ref. No.	Methodologies	Limitations / Demerits of Proposed Work	Advantages of Proposed Work	Accuracy (%)
[20]	In this study, a machine-learning pipeline was proposed to predict knee joint space narrowing (JSN) in KOA patients. The proposed methodology, which is based on multidisciplinary data from the osteoarthritis initiative (OAI) database, employs: (i) a clustering process to identify groups of people with progressing and non-progressing JSN; (ii) a robust feature selection process consisting of a filter, wrapper and embedded techniques that identify the most informative risk factors that contribute to JSN prediction.	Image-based deep learning algorithms to extract morphological knee features is not included.	In this study, the problem of JSN prediction in knee osteoarthritis patients was investigated. The main finding was that a combination of heterogeneous features coming from almost all feature categories is needed to maximize the performance of the predictive models.	78.3
[21]	The study introduces a new approach to quantify knee osteoarthritis (OA) severity using radiographic (X-ray) images. The new approach combines pre-processing, Convolutional Neural Network (CNN) as a feature extraction method, followed by Long Short-Term Memory (LSTM) as a classification method [2].	The segmentation task of X-ray Images had not been implemented which would have helped to accurately predict the pixel label for JSN with a limited amount of training data.	Obtained results are more accurate with a mean accuracy of 75.28%, and cross-entropy of 0.09, which shows that it outperforms the previous deep learning models implemented for similar issues.	75.28
[22]	A Support Vector Machine (SVM) based to predict the major OA risk factors and serum levels of adipokines/related inflammatory factors at the baseline for early prediction of at-risk knee OA patient structural progresses over time [3].	One of the major limitations of the SVM based model was that it was developed using the OAI cohort in which participants are at a mild-moderate stage of the disease and that the reproducibility analysis was performed with OA patients with more disease severity but mimicking clinical routine.	A Novel Approach, which was built for predicting knee OA structural signs of progress. Using this automated ML patient- and gender-based model, early prediction of knee structural OA progression can be performed with high accuracy using only three baseline serum biomarkers and two risk factors	80
[23]	A grading method for detecting and estimating the geometric and texture features of synovium thickening and bone erosion was proposed. Unlike previous studies in this area, this work uses the metrics and texture features of the region of interest (ROI). The highlighted feature of metacarpophalangeal bone and the dark feature of the synovial thickening is extracted simultaneously by the segmented method based on the Gaussian scale space. The segmented results are analyzed to extract three quantitative geometric parameters, which are combined with grey-level co-occurrence matrix (GLCM) statistic texture features to describe the ultrasonic Image of metacarpophalangeal RA. To obtain the preferable ability of classification, we applied a support vector machine (SVM) and various feature descriptors, including GLCM, local binary patterns (LBP), and GLCM C LBP, to grade the ultrasonic image of metacarpophalangeal RA [4].	More information should have been extracted to improve the performance of classification on metacarpophalangeal RA ultrasonic images.	This methodology points to a significant grading of metacarpophalangeal RA ultrasound images without medical expert analysis or blood sample analysis, such as detecting C-reactive protein, measuring erythrocyte sedimentation rate and testing rheumatoid factor.	92.50
[24]	Three ensemble algorithms, like SVM, Ada-boosting, and random sub-space, were used in this Investigation for the prediction of Rheumatoid arthritis (RA). These ensemble classifiers use k-NN and Random forest for baseline measurements of the classifier [5].	Although the use of the ensemble model was there it missed two things 1. Use of Neural networks as advanced techniques and 2. Use of image dataset for more accurate prediction of RA.	An ensemble approach for prediction of RA using a real-time dataset which gives greater accuracy using SVM and ADA-boosting and RSS with the use of baseline classifiers like RF and k-NN	90.50

.

Nonetheless, most of the studies are restricted to (i) a subset of patients, such as overweight and obese women, (ii) a single data modality, and most frequently on image-based algorithms utilizing MRI or X-ray images, (iii) questionnaires that lead to biased results due to the objectivity of patients, such as pain grade, and (iv) the accuracy of the model applied is not encouraging.

Therefore, we identified these holes in the prior literature, used image augmentation methods prior to the CNN design, and maximized the accuracy to 91.03%, which will aid in the identification of the KOA with ever-increasing precision.

3. Methodologies

3.1. System Architecture

Figure 3 represents the overall system architecture of our proposed model.

As mentioned in above Figure 3, we applied the CNN Inception Net V2 model for a knee X-ray image classification with severity grading. For this, first, we downloaded the dataset. The dataset contains the knee X-ray images. The pre-processing steps had been applied to the collected dataset. As the dataset is in image format, we applied some image enhancement techniques using an OpenCV library. The steps contain first the segmentation of images so that we can get the exact area of the knee for severity grading used in the CNN model. Then, we applied contour detection for having the edge detection of the target segmented area of the knee. We further used the image enhancement technique to improve the quality of images which will lead to a good feature extraction mechanism for the CNN Model. After applying the CNN Inception Net V2 model, we then classified the knee OA with 5 severity grading as the final results. All the above steps of system architecture are further explained in the following sections in detail.

3.2. Dataset Description

It is the goal of an Osteoarthritis Initiative (OAI) to identify biomarkers for the onset and development of knee osteoarthritis (OA) through a multi-center, long-term, prospective study. These pictures of the knee are utilized for assessment. It is possible to grade knee KL using the X-ray data in this collection. OAI (https://oai.epi-ucsf.org/datarelease/ (accessed on 31 January 2019)) is used to arrange the data. The following are the descriptions of each Grade:

• Grade 0: knee appears to be in good health in this photograph.
• Grade 1 (Unlikely): Possible osteophytic lipping with joint narrowing.
• Grade 2 (Minimal): Definite presence of osteophytes and possible joint space narrowing.
• Grade 3 (Moderate): Osteophytes, joint space narrowing, and mild sclerosis are all signs of disease.
• Grade 4 (severe): Osteophytes, joint constriction, and extensive sclerosis are all present in patients with disease.

3.3. Dataset Pre-Processing

All raw screened X-rays have their resolution scaled back to maintain consistency. Using a pixel density of 0.14 mm/pixel, we aim to get as near as possible to the physical resolution of all photographs. All processed photographs are then trimmed to a height and width of 2048 and 2560 pixels, respectively, to ensure uniformity. Since we can anticipate the KL grades of both knee joints from a single X-ray picture, all images with available KL ratings for both knee joints are preserved. After pre-processing and filtering, 4130 X-ray pictures of 8260 knee joints were still present. We used a 7:1:2 split of knee X-ray images for validation, training, and testing as part of our work. Depending on the KL grade of the left knee joint in the X-ray image, this division is carried out grade by grade. to guarantee that the distribution of grades across validation, training, and testing sets is consistent. This knee joint test set, after the split, consists of 1656 knee joints and 828 X-ray pictures, including 639 grade 0 s, 296 first grades, 447 second grades, 223, and 51 third grades.

3.4. Image Enhancement Technique

The original images downloaded from the source given above contains noise as it is the knee X-ray images obtained from the X-ray machine available at the hospital. Normally, there are chances that these images contain noise. Hence, we used the OpenCV library for image enhancement techniques, which will further enhance the results of classification using a neural network. Image enhancement techniques, e.g., contrast enhancement, contour detection using canny edge detection methods, image blur, and image sharpening, were attempted for research purposes. When we tried the above combination, we achieved improved results with the image sharping process.

Anterior knee joint view (ROI) was taken from the original X-ray images at baseline, which have previously been cropped. They’re 224 × 224 pixels in size. Images are split into train, test, and validation data. However, for our research, this split must match the ratio (80:20) in the other models using clinical data according to the division of clinical data For this reason, we re-arrange all images using a high-level file operation, shutil library in Python.

Image sharpening is used to decrease noise and equalize histograms in a customizable function first. These approaches include rotating photos 15 degrees, zooming to the nearest 0.1, shear mapping images, and turning half of the images horizontally. Pictured in Figure 4 is an example of what a picture after using the augmentation approach looks like.

Segmentation

To classify images of human knee OA, we employ the most recent unsupervised medical image segmentation technique based on the computation of the two-dimensional (2D) local center of mass (LCM) proposed by [22,24].

Step 1: Computation of CM (Centre of MASS)

The term "center of mass" refers to the value calculated from a cluster of pixels or a section of a curve based on their average intensity. The intensity value is used to determine the weights. The cluster’s centroid will be its central, most intense pixel. The formula for N-by-N intensity samples in a one-dimensional picture is f: R, where: = 1,..., N. Each pixel’s local center of mass (CM) is determined, and then pixels are clustered together into non-contiguous clusters.: C: Ω → ℝ.

Centre of mass is calculated as:

$$C_n=\raisebox{1ex}{${{\displaystyle \sum }}_{m=1}^N{W}_{m,n}m$}\!\left/ \!\raisebox{-1ex}{${{\displaystyle \sum }}_{m=1}^N{W}_{m,n}$}\right.$$

(1)

Step 2: Weight Calculation

An image’s intensity can be used to calculate the weight parameter required by Equation 1. A positive value, Wm,n is modest when the n-th and m-th pixels are in different regions and high when they are in the same region. The pixels from the same area should all be placed in the same cluster, as indicated by their shared center of mass C. Therefore, the evidence contained in the whole region is used to cluster the signal, rather than just the pixels immediately adjacent to it. The total cost of computing C (Cn|n ) is (N2).

Step 3: Initialization of Parameters

In order to perform the segmentation using the LCM method, some initial parameters must be determined, such as the iteration limit for LCM, which is set to 2000, and the computational parameters, which include and p, which are fixed based on the optimal results achieved via trial and error. A built-in Sobel operator [25] in the Watershed method is used to segment the first edges.

Step 4: Computing LCM

Images are broken down into groups according to their center of gravity. The iterative procedure is carried out until the specified goals have been met.

The whole segmentation procedure used in this study is depicted in the flowchart in Figure 5. The procedure begins with the import of knee X-ray pictures, followed by an initial quality screening, pre-processing of an image for noise removal, setup of parameters for the segmentation process, computing LCMs, and lastly segmenting the region of image based on CMs.

3.5. CNN Architecture

The Inception structure is combined with the Residual connection to generate Inception-Resnet-v2 [25,26]. Multiple convolutional filters are integrated there in the Inception-ResNet block via residual connections. The training time can be reduced by using residual connections instead of deep structures, which eliminates the problem of deterioration. Inception-Resnet-v2’s network architecture is seen in Figure 6.

In ImageNet, which is used to train CNN models, there are even more than 1 million natural photographs spread among 1000 different categories. A transfer learning technique may be used in [27], as the deep CNNs are sufficiently intelligent to learn general picture properties that apply to other image datasets without having to start from scratch. A single CNN’s transfer learning structure is shown in Figure 7. The last two layers in this design are classification layers with fully connected layers, while the pretrained networks serve as feature extraction for generic image features. This system is known as a single transfer learning network. Figure 6 shows a 1536-dimensional feature retrieved from our Inception-ResNet V2 model’s last fully connected layer.

Concatenating features taken from CNNs, we get such a 5632-dimensional feature vector. A diagram of the suggested feature concatenation technique is shown in Figure 8.

Features are combined and transmitted to two linked classification layers. In contrast to the network topologies, we add an additional hidden layer to the classification design. Thus, we are able to better utilize our network’s capacity to learn as well as adapt our CNNs’ general properties to particular microscopic image data. In [27], fine-tuning the pre-trained network is recommended. Two classification layers are sent to the concatenated characteristics. In contrast to network topologies, the categorization design includes an additional hidden layer. So, we may make use of our network’s ability to learn and adjust our CNNs’ general qualities for specific microscopic picture data. Making adjustments to the network that has already been trained has been advocated.

Based on concatenation of features, we have designed two-layer, fully connected transfer learning networks that may be used to microscopic image data. As may be seen in Figure 8, the proposed network’s whole design is shown.

3.6. Evaluation Parameters

On a scale of precision, recall, sensitivity, and level of accuracy, classifiers were judged. False negatives are possible. The system can create accurate predictions that are measured by the accuracy of its performance.

$$Accuracy=\frac{Correct Prediction}{Total Prediction}\ast 100$$

(2)

System sensitivity is a performance statistic that assesses a system’s capacity to accurately forecast favorable outcomes.

$$Sensitivity=\frac{True Positive}{True Positive+False Negative}\ast 100$$

(3)

Performance parameter Specificity assesses the system’s ability to correctly forecast negative outcomes.

$$Specificity=\frac{True Positive}{True Negative+False Positive}\ast 100$$

(4)

Precision refers to a system’s ability to produce just the most relevant data.

$$Precision=\frac{True Positive}{True Positive+False Positive}\ast 100$$

(5)

F-Measure utilizes the harmonic mean to combine accuracy and sensitivity measurements.

$$F-Measure=2\ast \frac{Sensitivity+Precision}{Sensitivity\ast Precision}\ast 100$$

(6)

4. Results and Discussion

Using the suggested CNN architecture, we’ll examine the outcomes of our experiments to measure the severity of knee OA.

4.1. Experimental Setup

There are two sets of photos in the dataset, one for the training and validation set, which contains 90% of the images, and the other for the testing set (10% of the images). Moreover, 50% of the photos are utilized for network testing, and the other 50% are used for validation in the testing and validation set. An early halt to avoid overfitting is used in the grid search for hyperparameter optimization. This criterion is dependent on the validation performance. If no additional progress is made after 500 iterations, the training will be discontinued.

An initial distribution with such a zero-mean Gaussian distribution and a standard deviation of 0.001 is used to generate the weights of the network. Using an exponential decay factor, the learning rate is modified

$$adaptive learning rate=learning rate\ast decay rat{e}^{\left(\frac{step}{decay step}\right)}$$

(7)

if the decay time is set to 1000. The trials begin with 12 epochs, then rise to 27 epochs, and the averaged testing results are also used to compare the outcomes. All three datasets are randomized for each test. Experiments were conducted using the Google Colab platform equipped with a graphics processing unit (GPU).

4.2. Experimental Results

The early phase involved applying the suggested CNN Inception ResNet v2 model to the KOA original dataset. As discussed earlier, after 27 epochs were executed, then the training accuracy was found as 78.65%, validation accuracy was 61.45%, and testing accuracy was found 72.03%. The confusion matrix is shown in Figure 9.

Table 2. Confusion Matrix of Model.

	Precision	Recall	F1-Score	Support
Minimal	0.88	0.81	0.85	29
Healthy	0.59	0.86	0.70	23
Moderate	0.55	0.43	0.48	15
Doubtful	1.00	0.42	0.59	31
Severe	0.78	1.00	0.88	8
Accuracy	-	-	0.72	-

shows the classification report of the model executed on the original KOA dataset.

In the second stage, the proposed CNN Inception ResNet v2 model was applied to an enhanced image obtained using image sharpening techniques discussed in Section 2. When the same Inception ResNet v2 model was applied to this new dataset of enhanced images, then training accuracy was found as 95.35%, validation accuracy was 74.58%, and testing accuracy was 91.03%. The confusion matrix is shown in Figure 10 and the classification report analysis in

Table 3. Classification Report on Image Enhancement Model.

	Precision	Recall	F1-Score	Support
Minimal	0.90	0.87	0.89	22
Healthy	0.78	0.91	0.90	19
Moderate	0.88	0.85	0.91	6
Doubtful	1.00	0.89	0.92	5
Severe	0.89	1.00	0.88	7
Accuracy	-	-	91.03	-

.

Figure 11a,b represents the accuracy and loss curves of the model before image enhancement, and Figure 12a,b represents the accuracy and loss curves of the model after image enhancement techniques.

As shown in Figure 11a, the training loss and validation loss is approximately 12%, training accuracy is 62%, and validation accuracy is 46%. These results are drawn on our model with the original images of KOA datasets. The testing accuracy was found as 72%.

As we had applied the image sharpening techniques for image enhancement and then applied the same model on enhancement images, we found the training loss to be around 8% and accuracy of training was 87.5%, validation accuracy was 80%, and testing accuracy was 91.03%. Figure 12a,b represents the same respectively.

Classification techniques for knee OA severity and their performance matrices are shown in

Table 4. Comparative Analysis of Proposed Work.

Method	Training Time (Seconds)	Testing Time (Seconds)	Accuracy (%)
VGG-16	760	48.15	70.73
RestNet-50	974.67	16.93	67.37
DenseNet-121	968	19.37	67.63
Antony et.al. [14]	423.67	15.74	63.4
Tiulpan et.al. [22]	150.67	3.8	67.71
Lim et.al. [21]	226.67	14.63	71.97
Chen et.al. [23]	364.67	25.43	69.7
Proposed Method	548.1	37.96	91.02

. We compared our results with those of prior studies that used deep learning to categorize the severity of knee OA and used VGGNet, ResNet, and DenseNet as the CNN architectures. In the past, researchers have only had a moderate degree of success when attempting to categorize the severity of knee OA, with the maximum reported accuracy being 71.97%. On the other hand, our approach using the image enhancement methodology yielded a comparable result with less data and achieved an accuracy of 91.03% when using CNN Inception ResNet v2. [26].

According to the comparative analysis of the contribution of our research work, after pre-processing, image segmentation, and image sharpening of the KOA dataset, the results have increased from 72% to 91.03% in terms of overall accuracy using the Inception ResNet V2 model.

From the comparison made among the proposed image enhancement methodologies using the Image ResNet V2 model, we conclude that:

• Images are processed by well before Images ResNet V2 CNNs for a specific dataset. In turn, this leads to differing skills for recording the small distinctions between KOA severity classifications. Thus, the applied CNN model shows varied performance in the different severity grades of KOA.
• The image enhancement technique (image sharpening) improved the classification performance of five-class KOA severity grade, which helps to overcome the limitations of the earlier implementation of CNN with less accuracy and lower performance and produces robust and superior performance in comparison with earlier work.

5. Conclusion and Future Scope

Predicting the overall depth of a knee osteoarthritis [28] using an image-enhanced system trained on the Inception ResNet V2 model is the focus of this study. The Kellgren–Lawrence scale was used to train pre-trained algorithms to provide accurate predictions about the severity of the disease. The transfer learning model ResNet V2 have been used to do the same job using X-ray pictures of the knee. What sets this work apart from others in this field is the use of sharpened photos from the KOA dataset, increasing the accuracy from 72% to 91.03% in this study. For the purpose of this study, it was determined whether or not the suggested fusion system in image enhancement could outperform the single component of deep learning. KOA severity was more accurately predicted by deep learning [29] fed with enhanced X-ray picture sharpness. However, combining patient data enables the identification of elements that influence illness development, which would be essential for medical professionals to provide an accurate diagnosis [30]. However, there are certain drawbacks to take into account, even though the suggested strategy helps to improve illness prediction. The models were run on a single dataset containing all of the data collected from the participants in the OAI research in the United States. In addition, the data used for this study are only derived from the study’s baseline [14]. In future, the author intends to work on the real-time dataset with precise severity grading.