A Comprehensive Review on the Diagnosis of Knee Injury by Deep Learning-Based Magnetic Resonance Imaging

The continual improvement in the field of medical diagnosis has led to the monopoly of using deep learning (DL)-based magnetic resonance imaging (MRI) for the diagnosis of knee injury related to meniscal injury, ligament injury including the cruciate ligaments, collateral ligaments and medial patella-femoral ligament, and cartilage injury. The present systematic review was done by PubMed and Directory of Open Access Journals (DOAJ), wherein we finalised 24 studies conducted on the accuracy of DL MRI studies for knee injury identification. The studies showed an accuracy of 72.5% to 100% indicating that DL MRI holds an equivalent performance as humans in decision-making and management of knee injuries. This further opens up future exploration for improving MRI-based diagnosis keeping in mind the limitations of verification bias and data imbalance in ground truth subjectivity.


Introduction And Background
The knee joint, one of the large complex joints, has been the topic of most discussions in view of the injuries to various anatomical structures -ligaments (anterior cruciate ligament, posterior cruciate ligament, medial collateral ligament and lateral collateral ligament), meniscus (medial or lateral) and cartilage [1,2].Knee injuries are reported to affect nearly 244,000 people annually, according to an analysis that included National Health Fund (NHF) data from 2016-2019 [1].The prevalence of knee pain was reported as 21.4% [3].In an Indian study, including 517 patients who underwent primary anterior cruciate ligament reconstruction (ACLR), 70% had a meniscal injury and 50% had chondral damage [4].Acute knee injury is generally caused because of direct trauma, or due to excess tension, sudden twists, collisions, awkward movements, falls, excessive force, and overuse of joints [5].Shoe wear, training surface conditions, and training regimen are the extrinsic factors, whereas ligamentous laxity, muscle weakness, reduced muscle flexibility, and foot shape are the intrinsic factors [6].The possible effects of knee injuries include tendinopathies and structural muscle injuries of the lower limb [7,8].Early detection of ruptured ligament, meniscal tears, as well as cartilage lesions and consequent treatment, are necessary for management, which can also delay the onset of knee osteoarthritis following a knee injury [9].
For the diagnosis, arthroscopy remains the gold-standard modality.However, its use over time has been restricted and overpowered by newer non-invasive investigations like magnetic resonance imaging (MRI) [10].However, diagnosis of knee injury by MRI can be difficult, with clinicians' experience being crucial in image interpretation and the best management of knee injuries is sometimes hampered by limitations of the human mind in interpreting the imaging reports by factors like a distraction, workload, subjectivity, image quality and knowledge [11].Furthermore, clinical-diagnostic differences between orthopaedic surgeons and non-musculoskeletal radiologists are frequently seen in day-to-day clinical practice [12].Considering the rising cases of knee injuries and sports injuries and the above-mentioned factors, the use of artificial intelligence (AI) has become rampant in medical practice, whereby certain diagnostic algorithms are created based on digital imaging data collection and interpolation.AI involves a technique that makes it possible for computers to complement human intelligence.The growth of AI is specifically being driven by deep learning (DL), which falls under the category of machine learning (ML) algorithms.There have been several documented uses of DL in image interpretation, including classifying skin carcinoma, detecting lung nodules, mammography cancer, and diabetic retinopathy.AI-powered technologies are anticipated to change the medical field as they increase the diagnostic accuracy of several diagnostic and therapeutic techniques [13].A number of early DL investigations have shown improved performance over conventional ML processes and are even found to be superior to radiologists in the diagnosis of knee injuries [14].The limitations of prior conducted similar systematic reviews were that they included other knee injuries like bony fractures [15] or failed to address the limitation and strengths of various AI in deciding the performance of the system [16].Taking into account the emerging AI technology as well as increasing research in similar fields, in this systematic review we conducted a comprehensive analysis of the literature, encompassing all DL-based methodologies which are employed in knee injury diagnosis.The present systematic review aimed to identify the latest studies that evaluated the role of DL in MRI-based knee injury diagnosis.The main emphasis was laid on research that assessed knee ligamentous tear, meniscal tears or cartilaginous lesions.

Literature Search
We conducted a thorough search of studies as per the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, wherein two primary databases, PubMed and Directory of Open Access Journals (DOAJ), were reviewed by two primary authors and the data was extracted by the third author.An electronic data search was conducted on the databases of PubMed and DOAJ over a period of the last 10 years of published studies with specified dates from January 2013 till January 2023.For the database search, Medical Subject Headings (MeSH) keywords were used as per the PubMed dictionary separated by Boolean expression: "knee"[MeSH Terms] OR "knee joint"[MeSH Terms] AND "magnetic resonance imaging" [MeSH Terms] AND "machine learning"[MeSH Terms] in "All fields" as tag terms.The text words used in DOAJ were the same as MeSH terms used in PubMed without any additional filters.The articles were found to be eligible based on the title and the abstract.The full text of the articles was studied only after verification of their abstract and the title and the inclusion criteria (Figure 1).

Inclusion Criteria
Only full texts and papers were chosen for the study.English articles were included which were published over the last 10 years from January 2013 till January 2023.Studies that diagnosed knee injury based on the MRI deep learning AI-based algorithms and studies that mentioned the accuracy of the performance of AI algorithms were included.

Exclusion Criteria
Exclusion criteria were articles published before January 2013, articles dealing with injuries other than knee injuries, studies published on animals, studies other than original articles, i.e. review articles, systematic research or meta-analysis, studies written in a language other than English, and editorial commentaries and book chapters.

Data Extraction
After downloading all the studies, the data was extracted and put in a table in Microsoft Spreadsheet.The information extracted from each article included the first name of the author, the year of publication, the description of data, the learning algorithm used, the sample size, the type of injury, and the accuracy of the DL model for diagnosis of a knee injury.

Assessment of Risk of Bias
The assessment of the risk of bias was done by using the ROBINS-I tool which consists of seven parameters that include selection bias, measurement bias, attrition bias, confounding bias, performance bias, and outcome reporting bias.A total score categorized as low, moderate, and high was classified, wherein the presence of none of these confounding factors or one of these confounding factors led to the classification of low and two or more up to five led to the classification of moderate and five or more up to seven led to the classification of high risk of bias as for the eligible studies is shown in Table 1 [17].

Search Results
Based on our search, we identified 203 articles in PubMed, 40 articles in DOAJ, and two articles from other references of the selected articles.Among them, there were 20 duplicate articles, which were excluded, and 225 articles were screened by two primary authors.Among them, 188 were excluded based on reasons that fell into the exclusion criteria and 37 full-text articles were downloaded.They were thoroughly screened, and 24 out of them were selected, and 13 were excluded since they did not have sufficient data to qualify for the systematic review.Finally, 24 studies were included in the systematic review, the characteristics of which are shown in  The countries where the included studies were conducted were France [22,32,34], Switzerland [25,28], Ireland [29], USA [13,20,24,26,30,31,33,35], Turkey [36] and Asia [18,19,21,23,27,[37][38][39][40].The study period in those studies varied from five years to 18 years.

Accuracy of the artificial intelligence model
The overall accuracy of the AI model was 72.5% to 100% for knee injuries.

Discussion
The present systematic review holds significance as it summarized all the studies that used deep learning MRI models for the diagnosis of knee injuries.Deep learning-based MRI is basically a part of artificial intelligence which is expanded in various domains in the entire world [41].Performance and accuracy as seen among the studies ranged from 72.5% to 100% which is effective enough for expanding the medical experts' knowledge and diagnostic skills for knee injuries.Our findings were in line with a previous systematic review conducted by Siouras et al. [41], who also quoted diagnostic accuracy of 72.5-100%, but the review was conducted on 22 studies.Another systematic review was conducted by Kunze et al. [16] including 11 studies, among which five evaluated ACL tears, five assessed meniscal tears, and one study assessed both where the area under the curve (AUC) for detecting ACL tear was in the range of 0.895 to 0.980 and for meniscus tear was 0.847 to 0.910.The use of deep learning has come frequently into practice since there is no gold standard scoring system to diagnose knee injuries.Artificial intelligence complements the human mind in a machine-operated way where the probability of diagnosing the injured knee becomes higher.This has been mainly based on the data augmentation and data acquiring for ACL, meniscal and cartilage injuries.Moreover, this has become possible through various image transformations which include shifting, and flipping rotations, thereby expanding the dataset and improving the performance of DL-based learning and diagnosis [41].We noticed that studies reported a difference in the performance of deep learning MRI-based diagnosis accuracy and this variability in the performance of different studies might be because of the region of interest that may appear slightly different within an image and because of the different ratios and sizes.The lowest accuracy of 72.5% was reported by Roblot et al. [34] and Bien et al. [13], which was primarily on the diagnosis of meniscus tears using convolutional neural network, while the highest accuracy was reported by Mazlan et al. [37] of 100%, which was specifically on ACL tears with the AI model of support vector machine (SVM) being used which gives high data gap between injury data (actual data) and non-injury data.This shows that variability in performance can be due to data imbalance whereby patients have different grades of knee injuries, and the application of a uniform algorithm may not be as effective.So, multiple datasets are needed whereby deep learning can be effectively improved by expanding the MRI protocol thereby allowing it to perform in equivalence to the human mind -stressing the role of different AI models to be used and evaluation of the wide type of knee injuries to improve on the accuracy.The present systematic review holds strength in this regard since it covers many regions allowing for data pooling of three categories of knee injuries -thereby representing the multiset of data with the elimination of the bias and determining the effective performance of deep learning-based MRI.

Limitations
The present study was a systematic review, but it has certain limitations.Firstly, the meta-analysis was not done.Secondly, studies do not assess the combined use of machine learning and human learning.Thirdly, the gold standard diagnostic arthroscopy was not done in all the studies, which may have restricted the clinical applicability of the findings of the present study.In view of these limitations, future studies are recommended to expand the datasets and test the accuracy of machine deep learning-based MRI for the detection of knee injuries, especially meniscal, ACL, and cartilage injuries, and compare them with the gold standard non-invasive tests so that their applicability can be put to use soon.This shall help in the future to cover up for the workload which is expanding by leaps and bounds, whereby radiological imaging data has been expanding, but the number of radiologists is not increasing proportionally.The decision-making process is also hampered on this account.Thus AI systems can be a boon to humans.This is also essential because medical imaging is very sensitive in nature whereby quality and resolution demand high performance with minimal differences in diagnosing specific knee injuries.Training and specialization of the human eye and mind may require long experience before one may come to terms with machine-based learning [12][13][14].

Future directions
The findings of the present study show the different applications of deep learning MRI used till now, and there is still so much scope to fully exploit the full potential of this data where newer algorithms can be laid down by individual hospitals for making trustworthy detection systems for knee injuries.This demands further AI expandability and a collaboration of medical and IT fields to reach precision medicine.

Conclusions
Deep-based learning methods have come a long way, showing performance accuracy of more than 75%, leading to significant use in clinical employment.However, there are so many algorithms to be laid down, and individual datasets need to be expanded by multiregional studies whereby deep-based MRI detection can become a norm in every hospital, thereby retaining the high-performance standards and yielding faster diagnosis and management.

FIGURE 1 :
FIGURE 1: Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Flowchart.DOAJ: Directory of Open Access Journals

TABLE 1 : Risk of Bias Assessment
L: Low, M: Moderate