Radiomics in application to diseases of the musculoskeletal system: a review

Full Text

INTRODUCTION

Medical imaging plays a crucial role in clinical practice by aiding doctors in making informed medical decisions. Currently, there are various imaging modalities available to clinicians, each involving different levels of complexity, resolution, and cost. These include ultrasound (US), plain X-ray (radiographs), computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MRI) [1].

Conventionally, medical images are processed by manually extracting qualitative visual data, such as the shape, size, and intensity of a region of interest. However, this approach may result in a late or incomplete diagnosis as pathological changes may become visible only in the late stages of a disease [2]. In addition, similar changes in the image may be observed for different pathological processes [3]. As such, there is a need for timely diagnosis and obtaining specific quantitative characteristics of radiological images for the appropriate interpretation of the obtained data [4]. At this point, radiomics can make a difference.

Radiomics employs mathematical algorithms to extract a large number of quantitative features (also known as radiomic features, RFs) from medical images. In doing so, radiomics can uncover image characteristics that are otherwise not visible to the naked eye, thereby increasing the amount of information that can be extracted from each medical image [5]. Initially, radiomics was applied to differentiate tumors in diverse biological tissues [6]. Tumor heterogeneity in terms of its microstructure and microenvironment is crucial for predicting the prognosis of a disease, planning therapy, and assessing the response to treatment. Therefore, radiomics can help personalize medical care by extracting useful information from medical images [7]. Furthermore, radiomics offers the potential to support drug development and evaluate treatment progress, which facilitates diagnosis and medical decision-making, making them faster and more accurate, while also tracking the disease progression.

The radiomics workflow aims to develop a classification model that discriminates pathology based on the relatively small number of RFs extracted from medical images. The radiomics workflow involves several steps, including image acquisition, image preprocessing, the region of interest selection, RF extraction, the selection of the relevant RFs, and construction and validation of a classification model (Fig. 1) [8, 9]. As the number of extracted RFs can exceed a thousand, machine learning (ML) algorithms are used for the analysis. The development of ML methods for radiomics feature analysis has demonstrated promising outcomes, especially in terms of the early detection of pathology and the prediction of disease prognosis [2]. Deep learning (DL), a subset of ML, has been successfully applied for automatic image segmentation (e.g., convolutional neural networks (CNNs) such as U-net and W-net) and for constructing classifier models based on deep neural networks (DNNs) [10, 11]. However, some studies have suggested that models based on deep features (i.e., pretrained models) do not always outperform the models based on analytically defined (radiomic) features [12].

 

Fig. 1. A schematic depiction of the radiomics workflow. MRI, Magnetic Resonance Imaging; CT, Computed Tomography; PET, Positron Emission Tomography; US, Ultrasound; VOI, Volume of Interest; LASSO, Least Absolute Shrinkage and Selection Operator; mRMR, Minimum Redundancy Maximum Relevance; ICC, Interclass Correlation Coefficient; SVM, Support Vector Machine; RF, Random Forest; XGBoost, eXtreme Gradient Boosting; ANN, Artificial Neural Network.

 

Despite its origins in oncology, radiomics is currently being applied in the study and diagnosis of a wide range of non-oncological diseases, including cardiovascular, neurological, respiratory, gastrointestinal, and hepatic diseases [13]. In recent years, several articles have been published on the use of radiomics in MSK diseases. This is particularly important for non-traumatic MSK diseases, which can be difficult to distinguish using conventional imaging methods, especially in the early stages when changes in biological tissues may not be visible to the naked eye [14]. Radiomics can overcome this challenge by providing a quick, accurate, and noninvasive approach to diagnose a pathology at an early stage [7]. By implementing radiomics in clinical practice and for early diagnostics, this technique has the potential to improve patient’s quality of life and reduce the socioeconomic burden on society [15]. Therefore, it is worth reviewing the results obtained so far in the radiomics field in light of potential prospects for research and applications.

This article provides an overview of recent research conducted on radiomics applications for the diagnosis of nononcological and nontraumatic diseases of the MSK and connective tissues.

ARTICLE SELECTION ALGORITHM

On April 3, 2023, the PubMed database was queried with the following parameters: Search in Title and Abstract; Keyword: radiomics; Excluding the following terms: Oncology, Cancer, Carcinoma, Glioma, Metastases, Tumor; Publication Language: English. PubMed query box:

(((((((radiomics[Title/Abstract]) AND (English[Language])) NOT (Oncology[Title/Abstract])) NOT (Cancer[Title/Abstract])) NOT (Carcinoma[Title/Abstract])) NOT (Glioma[Title/Abstract])) NOT (Metastases[Title/Abstract])) NOT (Tumor[Title/Abstract])

As a result, full bibliographic records of 1564 articles (including the following fields: PMID, Title, Authors, Citation, First Author, Journal/Book, Publication Year, Create Date, PMCID, NIHMS ID, DOI) were obtained and saved as a CSV file for further analyses.

Each article was then classified by title and abstract according to the ICD-10 by consensus of three authors (MP, EK, TK) based on the studied disease. No snowballing was applied. Overall, 29 groups were created with 26 groups labeled by Latin letters of the alphabet (A–Z), indicating a disease category according to the ICD-10. The “general” group included publications that did not focus on a specific disease or were related to radiomics methodology or algorithmic. The “review” group included review articles, and the “other” group included publications that could not be categorized into any of the abovementioned groups, primarily related to veterinary studies. Finally, 37 articles marked with the M category of the ICD-10 “Diseases of the musculoskeletal system and connective tissue” were selected for the review. Of these, 30 studies employed a retrospective design with sample sizes ranging from 36 to 7906 subjects, while the remaining 7 studies had a prospective design and included 25 to 731 subjects.

RADIOMICS APPLICATIONS IN DISEASES OF THE MUSKULOSKELETAL SYSTEM AND CONNECTIVE TISSUE

In recent years, an increasing number of studies have been published on radiomics applications for the detection of various diseases beyond oncology localized in different body parts [13]. Currently, the localization of the radiomics applications for MSK diseases includes the spine and large joints (knee, sacroiliac joint (SIJ), and hip); with less frequent localization in the temporomandibular joint (TMJ), shoulder tendon, Achilles tendon, and calf. The forearm, hand, and cervical vertebrae have not yet been used for radiomics analysis (Fig. 2). Currently, the most frequently studied topics among radiomics applications in MSK diseases include low bone mineral density (BMD), arthritis, and spondylitis (Table 1).

 

Fig. 2. Localization of the applications of radiomics in musculoskeletal system and connective tissue diseases: temporomandibular joint, shoulder tendon, hip, calf, Achilles tendon, spine, sacroiliac joint, knee; and the corresponding publications.

 

Table 1. An overview of the clinical applications of radiomics in non-oncological MSK diseases.

Setting	Applications	Modality used (number of applications)	Reference
Bone mineral loss/osteoporosis (n = 17)	• Diagnose, predict, and classify BMD • Prediction of residual back pain after vertebral augmentation • Detect vertebral fractures	• CT (n = 12, 71%), • DEXA (n = 5, 29%), • MRI (n = 4, 24%), radiographs (n = 1, 6%)	[16–18, 26–28, 30–34, 36–39, 50, 51]
Osteoarthritis (n = 6)	• Knee osteoarthritis detection	• MRI (n = 5), radiographs (n = 1)	[19–21, 53, 54, 56]
Spondyloarthritis and spondylitis (n = 4)	• Bone marrow edema • Low back pain • Sacroiliitis: inflammatory activity detection • Differentiating tuberculous spondylitis from pyogenic spondylitis	• MRI (n = 4, 100%)	[29, 42–44]
TMJ diseases (n = 3)	• Anterior disk displacement stage • Condition of the lateral pterygoid muscle in patients with and without rheumatoid arthritis • Prediction of the progress of TMJ damage in juvenile idiopathic arthritis	• MRI (n = 3, 100%)	[58–60]
Tendon and muscle diseases (n = 3)	• Achilles tendinopathy • Subacromial impingement syndrome • Facioscapulohumeral muscular dystrophy	• MRI (n = 1, 33%), • US (n = 2, 67%)	[22, 61, 63]
Other (n = 3)	• Lumbar disk herniation • Low back pain • Spondylitis myelopathy	• MRI (n = 3, 100%)	[23, 48, 49]
Note. CT, computed tomography; MRI, magnetic resonance imaging, DEXA, dual-energy X-ray absorptiometry; TMJ, temporomandibular joint; US, ultrasound; BMD, bone mineral density.

 

Manual image segmentation was applied to all studies, except for two that utilized semi-automatic segmentation tools [16, 17] and six that used automatic segmentation based on CNNs (U-net), toolboxes, or atlases [18–23] (Fig. 3a).

 

Fig. 3. (a) Distribution of the segmentation types used in the considered literature. (b) Top used methods to reduce dimensionality (t-test group includes its non-parametric analogs such as Wilcoxon and Mann–Whitney U-tests). LASSO, least absolute shrinkage and selection operator; mRMR, minimum redundancy maximum relevance; ICC, intraclass correlation coefficient; Corr coef, correlation coefficient (Pearson’s or Spearman’s); Log regr, logistic regression; RFE, recursive feature elimination; PCA, principal component analysis. (c) Top used classifiers. LR, logistic regression including Rad-score and Elastic Net; SVM, support Vector Machine; ANN, artificial neural network; KNN, K-nearest neighbors. (d) Top used MRI sequences.

 

Dimensionality reduction methods were adopted to decrease the number of the predictor variables from the value of over a thousand to a dozen to simplify the model, prevent model overfitting, shorten the training time, and improve the interpretation ability of the resulting model. A detailed description of the algorithms used is beyond the scope of this review; hence, we present a classification of the methods applied in the literature under consideration, as follows:

Dimensionality-reduction methods:

• Statistical tests:
• t-test
• Mann–Whitney U-test
• Wilcoxon test
• Feature selection:
• Embedded methods:

– Logistic Regression

– Least Absolute Shrinkage and Selection Operator (LASSO)

– Elastic Net

• Filter methods:

– Minimum Redundancy Maximum Relevance (mRMR)

– Intraclass Correlation Coefficient (ICC)

– Spearman/Pearson Correlation Coefficient

• Wrapper methods:

– Boruta

– Recursive Feature Elimination (RFE)

• Feature projection:

– Principal Component Analysis (PCA).

Among the considered publications, the absolute leader was LASSO regression (22 of 37), which employs L1 regularization to select important features (Fig. 3b).

The choice of classification model strongly depends on several conditions such as the type and amount of data, interpretability, and complexity of a model. Here, we indicate the most commonly used classification models without delving into extensive details: Logistic Regression (LR), Support Vector Machine (SVM), Artificial Neural Networks (ANN), and k-Nearest Neighbors (KNN) (Fig. 3c). Logistic regression is one of the simplest binary classifiers and also the most commonly applied one. In this review, we considered the radiomics score (Rad-score) and logistic regression as a single whole, as the Rad-score is generally built on the M remaining regression coefficients after feature selection (see Fig. 1):

$Rad-score=b_0+{\sum }_{i=1}^M{b}_i\times x_i$, (1)

where b₀, intercept, b_i, i-th regression coefficient, x_i, i-th selected feature.

The main imaging modalities used were CT and MRI, as supported by DEXA, especially in the case of BMD studies. Less frequently US and radiographs were used. The most frequently applied MRI sequences were T2 and T1, followed by STIR, SPAIR, and, finally, Proton Density (Fig. 3 d). MRI was predominantly applied to visualize the pathology of soft tissues, such as muscles, tendons, and joints.

Different types of predictive models were present in the studies considered below: clinical, radiomics, deep, clinical-radiomics, radiomics-deep, and clinical-radiomics-deep models (Fig. 4a). Mainly, they differ in the origin of the features (predictors) used in the model. For instance, a clinical model relies on demographic parameters such as age, sex, and weight, and/or the results of clinical examination such as BMD, alkaline phosphatase, and homocysteine levels. Radiomics models are based on features, such as intensity, shape, and texture, as extracted from an image by specially defined mathematical algorithms. Deep models include features extracted from an image via ANN. Finally, a combined model is a combination of the above-mentioned types. The most prevalent model types were radiomics and clinical-radiomics (Fig. 4 b): nearly 20 studies developed such types of models, whereas the models based on the deep features were the least common ones: only 6 studies implemented NNs in model development either way.

 

Fig. 4. a, Different input sources for the classification models and their types; b, number of developed classification models per type in the considered literature.

 

The receiver operator curve area under the curve (ROC AUC) is generally applied to demonstrate the model performance, where a value of 0.5 indicates an absolute random guess (50/50 chance) and a value of 1 indicates a perfect classification. In addition, a ROC AUC score >0.8 is considered good, while a score >0.9 is considered great [24].

PATHOLOGICAL NONONCOLOGICAL CHANGES IN THE SPINE AND SACROILIAC JOINTS

Bone Mineral Loss

Bone mineral loss (BML) is considered as the cause of osteoporosis (OP). Dual-energy X-ray absorptiometry (DEXA, DXA) is commonly used to measure the central skeletal BMD, particularly of the spine vertebrae and femoral region [25]. Yao et al. [26] established and validated a radiomics nomogram based on the fat-water imaging of dual-energy spectral CT images of T11-L2 vertebrae for diagnosing low BMD with a great AUC of 0.98. Dai et al. [27] showed another radiomics model (RM) based on abdominal CT that could predict the BMD of lumbar vertebrae. As such, the radiomics approach may be promising for classifying diseases associated with BMD loss based on the features extracted from DEXA. For instance, Rastegar et al. [28] developed radiomics predictive models to classify patients with OP or osteopenia and healthy individuals based on DEXA images of the L1-L4 vertebrae and three femoral regions. The highest AUC of 0.78 was achieved by the model detecting OP in the trochanteric region. Moreover, the authors identified one of the most significant RFs for each model: normal vs. OP, normal vs. osteopenia, normal vs. OP /osteopenia, and osteopenia vs. OP.

Osteoporosis and Osteoporotic Fractures

OP is a disease that weakens bones due to the reduced rate of bone creation compared with bone loss. OP can result in the development of severe complications in the form of fractures (such as vertebrae, proximal femur, and distal radius); therefore, OP screening is essential for timely treatment. However, OP is often asymptomatic and painless until the moment of fracture, and there may not be indications for undergoing DXA [29].

Several publications [16, 30–32] developed RMs to diagnose OP based on CT images of the lumbar vertebra. These studies revealed that radiomics analysis based on lumbar spine CT can be an effective approach for preoperative OP screening and may serve as an alternative to bone health screening. Currently, the procedure of radiomics analysis seems complicated; however, advances in automated segmentation could facilitate the integration of feature extraction and calculation into the software. However, this method cannot replace DEXA, which remains the standard examination recommended by the ISCD. Unlike in the abovementioned studies, Xue et al. [18] applied automatic segmentation of CT images of the lumbar vertebra by a NN for further OP diagnosis, achieving a maximum AUC of 0.994. Notably, He et al. [33] used three MRI sequences: T1, T2, and T1 + T2 to obtain images of the lumbar vertebra to discriminate 1) normal vs. osteopenia, 2) normal vs. OP, and 3) osteopenia vs. OP. Nearly all models performed well (AUC > 0.7), whereas the best model was normal against osteopenia at the T1 + T2 operation mode. Zhao et al. [34] employed a combination of modalities (CT + MRI) and manual segmentation to develop an automated pipeline for osteopenia and OP detection and OP prediction with an excellent performance of AUC = 0.925 and 0.899, respectively. Moreover, the manual segmentation of L1–L3 was used to train U-net for further automatic segmentation. Most of these studies (except those by He et al. [33] and Zhao et al. [34]) were based on routine preoperative CT scans of the lumbar spine that were used to construct RMs for future OP screening.

Osteoporotic fracture is a severe complication of OP that requires early identification and prevention [35]. Several studies have investigated the role of CT-based radiomics in predicting osteoporotic fractures. Particularly, Wang et al. [36] evaluated the performance of a developed RM for osteoporotic vertebral fractures (OVFs) prediction in a longitudinal analysis based on CT images achieving an AUC of >0.7. The research conducted by Ge et al. [37] established a radiomics score-based nomogram for the preoperative prediction of residual back pain in osteoporotic vertebral compression fracture patients based on X-ray, CT, and MRI images. Yang et al. [38] developed an RM aimed at distinguishing between acute and chronic OVF based on CT images. Biamonte et al. [39] developed a RM based on CT images of the lumbar spine to detect fragility vertebral fractures, achieving an AUC of 0.784.

Spondyloarthritis and Spondylitis

Spondyloarthritis (SpA) is a group of chronic inflammatory diseases of the spine, joints, and enthesis, which are characterized by common clinical, radiological, MRI, and genetic features [40, 41]. All radiomics-based studies of SpA considered below are based on MRI of the SIJ. Ye et al. [42] proposed a radiomics-clinical nomogram model for differentiating axial SpA and nonaxial SpA in low back pain with an excellent AUC performance of 0.9. A.P.M. Tenório et al. [43] reported that the application of the radiomics approach constitutes a potential noninvasive tool to aid in diagnosing sacroiliitis and SpA subclassifications. Zheng et al. [44] developed a radiomics-based method to evaluate bone marrow edema (BME) in patients with axial SpA, scoring a great AUC of 0.9. The authors considered RM as an alternative to the SPARCC (Spondyloarthritis Research Consortium of Canada) [45] assessment system for the quantitative assessment of BME in the SIJ with axial SpA. The Rad score is considered a highly reliable and objective quantitative assessment of SIJ BME, a promising assessment of the dynamics of changes after treatment. Finally, Tenorio et al. [29] discovered that RFs support the clinical assessment of SpA based on SPAIR and STIR MRI sequences. Moreover, the authors developed a radiomics-deep model for inflammatory sacroiliitis detection, which revealed diagnostic performance comparable to that of experienced radiologists.

Tuberculous spondylitis (TS) is one of the most frequent and severe extrapulmonary forms of osteoarticular tuberculosis. Late or incorrect treatment of TS can result in the development of various complications and consequences. In critical cases, damage to the spine can result in spinal deformity, dysfunction of the genitourinary system and intestines, neurological disorders, and disability. However, the onset of TS is non-specific: clinical symptoms, laboratory findings, and medical imaging are nonspecific [46]. Wu et al. [47] developed a clinical-radiomics nomogram (AUC 0.83) for differentiating TS from pyogenic spondylitis based on CT images of vertebrae and clinical risk factors.

Other Diseases of the Spine

Yu et al. [48] developed a nomogram (AUC 0.93) to provide conservative treatment efficacy in quantitatively providing a reference for clinicians to avoid inadequate and excessive treatments of lumbar disc herniation based on T2-weighted (T2W) MRI of the most prominent intervertebral disks.

Moreover, Song et al. [49] developed a clinical-radiomics nomogram (AUC 0.84) based on the axial T2 MRI of lumbar soft tissues to analyze changes in the lumbar fascia of patients with low back pain. Segmentation was performed at the midpoint of the axial T2W image of the L4–L5 disks, including the erector spinae and multifidus at this level and the part of the fat behind the muscle that contains the fascia. The developed model could discriminate early changes in the superficial and deep fascia that are otherwise difficult to monitor in subjects with low back pain.

The application of MRI of the cervical spine to predict postoperative neurological function, despite its widespread use, may not be satisfactory in patients with spondylotic myelopathy. Therefore, Zhang et al. [23] developed an RM (AUC 0.74) for predicting cervical spondylotic myelopathy (spinal cord compression) based on T2*W MR images of the maximum compressed level of the spine.

PATHOLOGICAL NONONCOLOGICAL CHANGES IN THE LARGE PERIPHERAL JOINTS

Transient osteoporosis

Transient osteoporosis of the hip (TOH) is a rare disease of unspecified etiology that is characterized by transient neurovascular disorders that trigger processes resulting in a local decrease in BMD, which increases the risk of low-energy or pathological fracture of the proximal femur. Particularly, this process is local in nature and topographically limited to the proximal femur, which distinguishes it from generalized OP, as characterized by the violation of the microarchitecture and a decrease in the mineral density throughout the skeleton. In addition, TOH is characterized by drug-free self-resolution of the process or a good drug response to conservative treatment, allowing differentiation from avascular necrosis of the femur, which requires surgical intervention [29]. MRI plays a primary role in the differential diagnosis of these diseases, albeit, in some cases, it is difficult to distinguish the nature of BME. Therefore, a misdiagnosis can have a crucial impact on the treatment strategizing and lead to unnecessary surgery. Klontzas et al. [50] used MRI images of the hip, radiomics, and ML for accurate differentiation between AVN and TOH. The final RM achieved performance similar to that of MSK radiologists with an AUC of 0.937 and significantly greater performance when compared to that of general radiologists. Next, S Kim et al. [51] proposed that RM combined with a DNN can help diagnose OP based on plain X-ray images (radiographs) of the hip with a high-performance AUC of 0.95. Deep features were extracted using two versions of the DL model. The cropped image model included the cropped region of the proximal femurs on both sides. The image model relies on radiographs of the entire pelvis and hip. This approach can serve as the basis for sorting patients at risk of OP for further diagnostic tests, including DXA. Lim et al. [17] aimed at the RM development for detecting femoral OP based on abdomen-pelvic CT using semiautomatic segmentation (region growing editing tool) AUC 0.96.

Osteoarthritis

Osteoarthritis (OA) is a heterogeneous group of diseases with various etiologies that share similar biological, morphological, and clinical manifestations and outcomes, which primarily affect all joint components such as the cartilages, subchondral bone, synovial membrane, ligaments, capsule, and periarticular muscles [29]. During OA formation, changes in the subchondral region are accompanied by cartilage damage. The treatment of OA, directed at the condition of the subchondral bone, slows disease progression and improves the patient’s quality of life [52]. Therefore, early sensitive diagnosis of bone changes in OA is essential.

All publications related to radiomics-based diagnostics of OA have used MRI images of a knee. For instance, Xue et al. [19] developed the RM of subchondral bone assessment for the diagnosis of knee OA (KOA). The RM showed greater diagnostic efficiency compared to that of a model based on trabecular morphological parameters with AUCs of 0.961. The radiomics model based on MRI and clinical data in the study by Lin et al. [53] exhibited a high predictive performance (AUC 0.78) in differentiating knee pain and detecting OA. Xie et al. [54] developed a new implementation of radiomics analysis for the cartilage and subchondral bone of the knee, both showing excellent AUCs of 0.982 and 0.939, respectively. The RM was superior in detecting knees predisposed to post-traumatic OA after anterior cruciate ligament reconstruction relative to the classic T2 value-based model.

In addition, some quantitative assessment methods exist for menisci and articular cartilage next to the semiquantitative approaches used for assessing structural tissue disorders in OA on MRI. Because the subchondral bone is involved in OA, bone image quantification can be crucial for the detection, monitoring, and prediction of OA. Hirvasniemi et al. [20] assessed the ability of semi-automatically extracted RFs from the tibial subchondral bone to distinguish between knees without and with OA using automatic atlas-based segmentation with a good performance (AUC 0.8).

The infrapatellar fat pad (IPFP) is a fatty tissue located below the patella that influences the formation of OA. Therefore, (semi)quantitative analysis of changes in the IPFP can reveal the onset, development, and KOA prognosis [55]. Yu et al. [21] investigated whether the IPFP RFs from MRI images, extracted using automatic 2.5D CNN U-net segmentation, had predictive value for incident radiographic KOA a year before its diagnosis. Clinical, radiomics, and combined (clinical-radiomics) models were constructed and compared. Both the clinical-radiomics and radiomics models performed better than the corresponding clinical models (AUC 0.702).

The only study in the literature to have used plain X-ray images to develop a RM was that by Li et al. [56] in relation to the diagnosis of KOA. The developed clinical-radiomics nomogram model, based on the anterior-posterior and lateral X-ray images and age, was efficient in the KOA diagnosis (AUC 0.843).

TEMPOROMANDIBULAR JOINT DISORDERS

Temporomandibular joint (TMJ) lesions (dysfunctions) are a common problem, affecting up to one-third of all adults [57]. MRI is a common method for diagnosing TMJ conditions and can assess the disc, especially in soft tissue pathologies. Conventional T1-weighted (T1W), T2-weighted (T2W), and proton-weighted pulse sequences are used in MRI examination of the TMJ. However, the early detection of inflammatory changes in the TMJ is challenged by its small size.

Orhan et al. [58] constructed a radiomics model to identify anterior disc displacement without or with reduction based on bilateral MRI images of condyles and discs, with the model achieving an AUC of 0.77. Muraoka et al. [59] analyzed the texture of the lateral pterygoid muscle to identify RFs, which discriminated rheumatoid arthritis from normal MR images. Ricardo et al. [60] evaluated biomarkers (i.e., RFs) capable of diagnosing juvenile idiopathic arthritis in the TMJ based on MRI images of the mandibular condyle and identified correlations of the extracted biomarkers with the patient’s age, gender, and disease onset age.

TENDON AND MUSCLE DISEASES

Achilles tendinopathy is most often the result of a sports injury. Currently, US diagnostics can detect the thickening of the Achilles tendon aponeurosis, a decrease in the echo signal, the thickening of the Achilles tendon, hyperplasia of blood vessels, increased echo from the fat pad, local calcification, and other changes. In addition, Wang et al. [61] analyzed US images using radiomics to detect changes in the tissues of the Achilles tendon and elucidated the connotative characteristics of Achilles tendinopathy at a high-performance AUC of 0.99. The US sensor was always placed perpendicular to the long axis of the tendon to avoid image anisotropy. Red, green, blue and grayscale channels were used to extract RFs from the Achilles tendon region.

Quantitative MRI (qMRI) allows quantitative tissue mapping for numerical measurements in line with anatomical ones [62]. G. Colelli et al. [22] evaluated the feasibility of using a conventional short-tau inversion recovery (STIR) sequence to obtain qMRI parameters (predict fat fraction and water T2) by introducing a radiomics workflow in patients with facioscapulohumeral muscular dystrophy. The acquisition volume was centered on the calf, with the last acquired slice located 6-cm proximally from the upper limit of the patella. An automatic segmentation based on DNN was applied for the Soleus, Medial, and lateral gastrnemius, anterior tibialis, extensor digitorum longus, and peroneus longus muscles [46]. The combination of a workflow based on two new STIR-based RFs and a KNN as a classifier is reportedly the best predictor (with the least mean absolute error) of the fat fraction and water relaxation time.

Moreover, Jiang et al. [63] developed an ultrasomic (US + radiomics) model [64] that helps to identify the stage of subacromial impingement syndrome (whether inflammation and irritation of the shoulder tendons), which may facilitate the preliminary screening of shoulder pain with an AUC of 0.789.

In summary, we identified that the scope of radiomics applications in MSK diseases is limited by injuries, several inflammatory diseases of the spine and peripheral joints, and potential BMD loss. All of the above-discussed studies have presented models with high predictive ability, as confirmed by high values of AUCs ranging from 0.7 to 0.99 (mean 0.87). We believe that these RMs can help in clinical decision-making in the early stages of MSK diseases, thereby facilitating the initiation of appropriate treatment, especially in cases where it is difficult to detect significant changes in the images.

ASPECTS OF RADIOMICS ANALYSIS

The universal nature of radiomics lies in the processing of any digital image (two- or three-dimensional) as numerical data. Consequently, all available clinical imaging modalities can be applied for radiomics analysis. Among the most commonly used modalities are three-dimensional MRI, PET, and CT, followed by the less frequently used two-dimensional X-ray imaging (radiography) and US imaging [63]. Some studies have used a combination of modalities, such as MRI with CT [33, 34, 38] to obtain detailed information. US images inherit an extra source of variation related to the detector alignment, relative to the studied object [61]. Therefore, the US was the least commonly used modality in radiomics studies under consideration. PET was not mentioned in any of the relevant studies.

Despite the advances of NNs in biomedical image segmentation [11, 65], radiomics applications in MSK diseases predominantly involve manual segmentation. This aspect can be attributed to the complexity of NN training, which warrants an extensive training dataset. Automatic segmentation speeds up the processing of large image datasets compared to manual segmentation. CNNs such as U-net or W-net can be applied for automatic image segmentation because their architecture is well-known and has already proven to be efficient in medical image segmentation [11]. Thus, the advent of ML opens up promising prospects for automatic segmentation [66].

Multi-source input models (such as clinical-radiomics and clinical-radiomics-deep) are more common, probably due to their better performance relative to that of single-source input (such as clinical only and radiomics only) models. This observation can be explained by the presence of additional independent information sources in the model. For example, the study by Kim et al. [51] compared different models and demonstrated that the model incorporating all features (i.e., clinical, textural, and deep) indicated the best classification performance. Nevertheless, DL models for digital image classification incorporating only ANNs can serve as an alternative to the radiomics approach. The limitations of the DL approach include the complicated interpretation of features and outcomes, manual segmentation, and the large dataset required for NN training [67].

Ultimately, the common radiomics model in MSK diseases looks as follows: “Radiomics or clinical-radiomics model based on manually segmented MRI or CT images that used LASSO for important feature selection and logistic regression as a classifier.”

Since its origin in 2012, radiomics has primarily focused on oncology. However, research on the nononcological applications of radiomics has steadily increased [13]. Among the nononcological diseases, the visualization of the pathology of the MSK is important, albeit the introduction of radiomics into this area has only begun. The prospective application of radiomics in rheumatologic diseases seems promising for an early diagnosis of soft-tissue changes, especially with the use of conventional single-energy CT images. For example, diverse types of microcrystalline arthropathies, particularly gouty arthritis, induce morphological changes in the soft tissues, which can be potentially detected by radiomics at an early stage.

FUTURE PROSPECTIVE

The implementation of radiomics-based tools in clinical practice requires an interdisciplinary approach that involves close cooperation among medical doctors, data analysts, and software developers. Radiomics can help in the early diagnosis and monitoring of the disease as well as in the prediction of the treatment efficiency and the risk of recurrence of several diseases, including MSK diseases. Although radiomics applications in MSK diseases lag behind other narrow fields such as cardiovascular, neurological diseases, COVID-19, and others in the application of radiomics [13], its future development seems promising.

Future research should probably extend more toward the use of plain X-ray images due to their wide availability in clinics. Moreover, the future development of radiomics will face several challenges, such as the creation of datasets of medical images acquired using different modalities. These datasets will serve as a basis to train CNNs for automatic segmentation as well as to construct various classifiers such as ML- and DL-based, with the latter especially requiring large datasets. Since collecting large medical datasets requires substantial time and effort, data augmentation methods would be necessary to increase the dataset size.

All considered studies have the potential to support a clinician in decision-making by noting small changes in the tissue of interest. However, among the limitations, one can indicate manual segmentation, small sample size, the lack of external validation, and a nontransparent description of the conducted research. All of these limitations strongly limit the reproducibility of the developed RMs and, consequently, the clinical implementation. Therefore, cross-validation of the developed models using external datasets (from other institutions, devices, and patients) is essential before its implementation in a clinical setting [68].

In the future, it would be advisable to create an open platform for sharing radiomics models, similar to the Open-WebUI¹ platform [69] for large language models. This step would allow researchers and clinicians to access pretrained models and use them in their clinical settings while providing a means for developers to receive feedback on the outcomes. Such a platform could greatly enhance the validation of RMs.

Considering the dynamic and progressive development of radiomics, newer tools for assessing the quality of radiomics studies are emerging. One of the recent tools is the Methodological Assessment of Radiomics (METRICS), which was developed by an international consortium of experts in the field and endorsed by the European Society for Medical Imaging² [69]. METRICS provides a well-designed framework for assessing the quality of radiomics research by using a flexible format that covers all methodological variations. METRICS should facilitate the promotion and consolidation of the highest quality studies in radiomics [70].

CONCLUSION

Radiomics applications are rapidly expanding in the area of nononcological diseases, particularly in MSK diseases. Clinical-radiomics models are the most common ones used in this field because of their better performance compared to single-source input models (clinical only and radiomics only). Implementing artificial intelligence methods in musculoskeletal radiomics is limited to single applications in segmentation and classification. However, the further extension of radiomics for the early diagnosis, monitoring, and treatment of rheumatological and orthopedic diseases and injuries appears promising.

ADDITIONAL INFORMATION

Funding source. This article was not supported by any external sources of funding.

Disclosure of interests. The authors declare that they have no relationships, activities or interests (personal, professional or financial) with third parties (commercial, non-commercial, private) whose interests may be affected by the content of the article, as well as no other relationships, activities or interests over the past three years that must be reported.

Authors’ contribution. M.O. Pleshkov: concept of the work, collection and analysis of literature data, writing the manuscript; M.A. Zamyshevskaya, I.V. Tolmachev: concept of the work, writing the manuscript; E.V. Kuchinskii, T.V. Kim: collection and analysis of literature data; X.Jin: concept of the work, writing and editing the manuscript; J. Zhang writing the manuscript; V.D. Zavadovskaya, M.A. Zorkaltsev, D.A. Pogonchenkova editing the manuscript; V.D. Udodov: concept of the work. Thereby, all authors provided approval of the version to be published and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

¹ Open-WebUI [Internet]. GitHub; 2022–2024. From: https://github.com/open-webui/open-webui Last accessed March 1, 2024.

² EuSoMII [Internet]. Vienna: Europen Society of Medical Imaging Informatics; 2020–2024.From: https://www.eusomii.org/ Last accessed March 1, 2024.

About the authors

Maksim O. Pleshkov

Siberian State Medical University

Author for correspondence.
Email: maksim.o.pleshkov@gmail.com
ORCID iD: 0000-0002-4131-0115
SPIN-code: 8625-0940
Russian Federation, Tomsk

Maria A. Zamyshevskaya

Siberian State Medical University

Email: zamyshevskayamari@mail.ru
ORCID iD: 0000-0001-7582-3843
SPIN-code: 4434-1179

MD, Cand. Sci. (Medicine)

Russian Federation, Tomsk

Egor V. Kuchinskii

Siberian State Medical University

Email: egorelsigich@gmail.com
ORCID iD: 0009-0002-5960-0935
Russian Federation, Tomsk

Xiance Jin

1st Affiliated Hospital of Wenzhou Medical University

Email: jinxc1979@hotmail.com
ORCID iD: 0000-0002-4117-5953
China, Wenzhou

Ji Zhang

1st Affiliated Hospital of Wenzhou Medical University

Email: jizhang1996@126.com
ORCID iD: 0000-0002-2718-6509
China, Wenzhou

Vera D. Zavadovskaya

Siberian State Medical University

Email: wdzav@mail.ru
ORCID iD: 0000-0001-6231-7650
SPIN-code: 7905-8363

MD, Dr. Sci. (Medicine)

Russian Federation, Tomsk

Maxim A. Zorkaltsev

Siberian State Medical University

Email: zorkaltsev@mail.ru
ORCID iD: 0000-0003-0025-2147
SPIN-code: 3769-8560

MD, Dr. Sci. (Medicine)

Russian Federation, Tomsk

Tkhe V. Kim

Siberian State Medical University

Email: Pavel.kim.08@mail.ru
ORCID iD: 0009-0002-9766-6986
SPIN-code: 7834-9024
Russian Federation, Tomsk

Daria A. Pogonchenkova

Siberian State Medical University

Email: azarova_d_a@mail.ru
ORCID iD: 0000-0002-5903-3662
SPIN-code: 4141-9068

MD, Cand. Sci. (Medicine)

Russian Federation, Tomsk

Vladimir D. Udodov

Siberian State Medical University

Email: linx86rus@gmail.com
ORCID iD: 0000-0002-1321-7861
SPIN-code: 3619-0496

MD, Cand. Sci. (Medicine)

Russian Federation, Tomsk

Ivan V. Tolmachev

Siberian State Medical University

Email: ivantolm@mail.ru
ORCID iD: 0000-0002-2888-5539
SPIN-code: 1074-1268

MD, Cand. Sci. (Medicine)

Russian Federation, Tomsk

References

Supplementary files