Sage Journals: Discover world-class research

Abstract

Background

Superior labrum anterior and posterior (SLAP) lesions are a common cause of shoulder pain and instability. Accurate diagnosis remains challenging in clinical practice. This study aims to develop and evaluate radiomics models and combined models integrating radiomics and clinical features for SLAP lesion detection.

Methods

This retrospective study included 149 patients who underwent shoulder arthroscopic surgery with preoperative shoulder magnetic resonance imaging (MRI) between 2019 and 2024. Regions of interest (ROIs) were manually delineated on MRI oblique coronal proton density-weighted fat-suppressed (PD FS) images, and radiomics features were subsequently extracted from these defined regions. Feature selection employed independent t-tests, Mann-Whitney U tests, Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Common machine learning models including Support Vector Machine (SVM), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM) were employed to construct diagnostic models based on radiomics features. A combined model integrating radiomics and clinical features was developed and visualized using nomograms.

Results

In the test cohort, the LightGBM-based radiomics model achieved optimal performance with the Area Under the Curve (AUC) of 0.867, sensitivity of 0.952, and specificity of 0.625. The combined model demonstrated enhanced diagnostic capability with AUC of 0.899, sensitivity of 0.762, and specificity of 0.917. Manual diagnosis of SLAP injury using MRI achieved an accuracy of 50.3%, with a sensitivity of 27.7%, specificity of 78.8%, and AUC of 0.619.

Conclusion

Machine learning models based on MRI radiomics features demonstrated superior diagnostic accuracy compared to traditional radiologist assessment for SLAP lesions. The combined model incorporating both radiomics and clinical features provides effective risk prediction for SLAP lesions.

Keywords

SLAP lesion radiomics magnetic resonance imaging machine learning

Introduction

Superior labrum anterior and posterior (SLAP) lesions of the glenoid labrum refer to the detachment of the superior labrum extending from anterior to posterior, potentially involving the attachment site of the long head of the biceps tendon.¹ Both the superior labrum and the biceps tendon anchor serve to enhance joint stability by acting as secondary stabilizers of the shoulder. SLAP lesions are considered a significant cause of shoulder pain and dysfunction. With the widespread adoption of arthroscopic techniques, an increasing number of SLAP lesions are being identified. Onyekwelu et al.² reported a 464% increase in arthroscopic SLAP repairs performed in hospitals in the northeastern United States between 2002 and 2010. Consequently, advancing diagnostic approaches for SLAP lesions has become essential for optimal patient management.

Physical examination and clinical presentations for SLAP lesions lack sufficient diagnostic accuracy and demonstrate considerable variability across studies.^3–9 Although MRI of the shoulder is currently the primary auxiliary diagnostic tool for SLAP lesions, studies have demonstrated that conventional magnetic resonance imaging (MRI) exhibits suboptimal diagnostic accuracy for SLAP lesions, with a sensitivity of 60% and specificity of 92.5%. This limitation may be attributed to the anatomical complexity of SLAP lesions.^10–12 Magnetic resonance arthrography (MRA) of the shoulder joint employs contrast enhancement by injecting a contrast agent into the glenohumeral joint space.¹³ Nevertheless, MRA has certain limitations, including its invasive nature, the risk of joint infection, and unsuitability for patients with contrast agent allergies. Therefore, MRA is typically reserved as a supplementary examination in cases where conventional MRI proves inconclusive, rather than being employed as a routine diagnostic measure.

In recent years, with the widespread application of machine learning and the rapid development of radiomics, researchers have begun to explore the application of radiomics in cartilage, ligaments, and tendons.^14,15 Oeding et al.¹⁶ developed an XGBoost-based machine learning model for subscapularis tear prediction using preoperative imaging, achieving excellent performance with an accuracy of 0.85. While Fei et al.¹⁷ established a radiomics-based model for rotator cuff tear diagnosis, achieving AUC of 0.989 and 0.979 in the training and validation cohorts, respectively. These models achieved favorable diagnostic performance.

Radiomics enables high-throughput quantitative extraction of high-dimensional imaging features from routine medical images. This approach allows for a comprehensive quantitative analysis of lesion heterogeneity, offering promising prospects in non-invasive preoperative diagnosis. To enable early detection and accurate diagnosis of SLAP lesions through more precise and non-invasive intelligent diagnostic approaches, the present study aims to establish a machine learning model based on clinical characteristics and radiomic features, and to evaluate its efficacy in diagnosing SLAP lesions. This research endeavors to provide valuable insights for the development of intelligent auxiliary tools in SLAP injury diagnosis.

Method

Methods and tools

Participants in the study

This study retrospectively collected clinical information and imaging data from patients who underwent shoulder arthroscopic surgery at our hospital from January 1, 2019, to December 31, 2024. Inclusion Criteria included: (1) Patients undergoing their first shoulder arthroscopic surgery at our hospital; (2) Preoperative shoulder MRI examination conducted at our hospital, including the following sequences: axial, oblique sagittal, and oblique coronal proton density-weighted fat-suppressed imaging (PD FS) and oblique coronal T1-weighted imaging (T1WI); (3) arthroscopically confirmed SLAP Type I or Type II lesions based on the Snyder classification system. Exclusion Criteria included: (1) Incomplete clinical data; (2) Missing MRI images, poor imaging quality, or incomplete MRI sequences; (3) Previous history of shoulder surgery; (4) arthroscopically confirmed Bankart lesions, shoulder dislocations, or glenohumeral instability. Figure 1 illustrates the flowchart depicting the process of study selection based on inclusion and exclusion criteria.

Figure 1.

Flow diagram of subject enrollment.

Collection of patient clinical information

We retrospectively collected clinical data from the hospital information system (HIS), including basic clinical information such as age, gender, history of shoulder injury, duration of illness, and imaging results from radiologists. A history of shoulder injury refers to any injury causing shoulder pain and/or limited mobility. The duration of illness was recorded in months. Two radiologists with experience in musculoskeletal imaging independently analyzed the shoulder MRI images, reaching a consensus after joint review in case of disagreement. They were blinded to any personal information or relevant clinical data of the patients during the assessment. All arthroscopic surgeries were performed by experienced surgeons from the orthopedics department of our hospital. Under arthroscopic assistance, through posterior and anterior/anterolateral approaches, the glenohumeral joint, long head of the biceps tendon, and biceps-labral complex were carefully examined. The condition of the superior labrum was observed and an intraoperative diagnosis was made based on the arthroscopic findings of SLAP lesions. If the superior labrum tear or detachment of the superior labrum from the glenoid surface at the biceps tendon anchor was found during surgery, it was recorded as a SLAP injury. In this study, only patients with SLAP lesions classified as types I to II according to the Snyder classification were included. Type I and Type II SLAP lesions were selected because they represent the most commonly encountered superior labral pathologies in clinical practice. Figure 2 shows images from intraoperative exploration. Simultaneously, the glenohumeral joint cavity and subacromial bursa were meticulously examined to assess the extent of rotator cuff pathology. The condition of the rotator cuff tendons was thoroughly documented and categorized into three distinct classifications: intact, partial-thickness tear, or full-thickness tear. Both isolated SLAP lesions and SLAP lesions with concomitant rotator cuff tears (partial- or full-thickness) were included, with rotator cuff tears analyzed as a clinical variable.

Figure 2.

Arthroscopic examination of the superior labrum of the shoulder joint.

Research methodology

This study is a retrospective analysis that has been approved by the hospital ethics committee, adhering to ethical standards. The primary research protocol encompasses the following processes: (1) Acquisition of MRI images (2) Delineation of regions of interest (3) Extraction and screening of radiomics features (4) Construction and evaluation of radiomics models (5) Assessment of traditional manual diagnostic efficacy (6) Development of an integrated diagnostic model combining radiomics features with clinical characteristics, and construction of a nomogram.

Magnetic resonance image collection and delineation of target regions

The magnetic resonance images utilized in this study were acquired using two 3.0 T MRI scanners (SIGNA Pioneer 3.0 T, GE Healthcare). All patients underwent non-contrast shoulder joint MRI examinations. For radiomics analysis, oblique coronal PD FS images were selected from each scan. The acquisition parameters were as follows: repetition time (TR) = 1800–2200 ms, echo time (TE) = 30–50 ms, slice thickness = 4 mm. All images were preserved with standard soft tissue settings and stored in Digital Imaging and Communications in Medicine (DICOM) format. ITK-snap software was employed to manually delineate the region of interest (ROI) of the superior labrum of the shoulder joint on the collected DICOM files.¹⁸ The superior labrum was contoured layer by layer along its margins on oblique coronal PD FS images. For each patient, 3–5 images were delineated from anterior to posterior, as illustrated in Figure 3. The areas delineated on these 3–5 images were saved as a single ROI file for each patient. Complete superior labrum regions were delineated in a total of 149 magnetic resonance images.

Figure 3.

Manually delineated region of interest on the superior labrum of the shoulder joint.

To evaluate the reproducibility and consistency of ROI delineation, intra-observer variability were assessed using the intraclass correlation coefficient (ICC). For intra-observer agreement evaluation, the same radiologist repeated the image segmentation in 30 patients who were randomly selected from the study cohort after an 8-weeks interval. ICC values were calculated based on a single rater, absolute agreement, 2-way random-effects model to assess the stability and reproducibility of radiomic features. Only cases with ICC values greater than 0.75 for ROI segmentation were included in the subsequent radiomics feature extraction and analysis to ensure reliable and reproducible results.

Extraction and screening of radiomics features

Radiomics features can be categorized into three distinct groups: morphological features, first-order features, and textural features. Morphological features delineate the three-dimensional shape characteristics of the contoured region. First-order features describe the first-order statistical distribution of voxel intensities within the delineated area. Textural features elucidate intensity patterns or spatial relationships among voxels within the contoured region. In this study, textural features were extracted using various methodologies, including the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM) approaches.¹⁹ All radiomics features were extracted using an internal feature analysis program implemented in pyradiomics (https://pyradiomics.readthedocs.io).²⁰ The radiomics feature extraction process adhered to the Image Biomarker Standardization Initiative (IBSI) guidelines.

All radiomics features were treated as continuous variables. For the features extracted from the images, we initially employed the Shapiro-Wilk test to assess their distribution, determining whether they followed a normal or non-normal distribution. Subsequently, we utilized the Mann-Whitney U test or independent samples t-test to evaluate significant differences in these features between groups. For normally distributed features, we employed the independent samples t-test for statistical testing and feature screening. For non-normally distributed features, we utilized the Mann-Whitney U test. Only radiomics features with a two-sided p-values below 0.05 between the two groups were retained. We conducted Pearson correlation tests on the features selected by the independent samples t-test or Mann-Whitney U test to calculate inter-feature correlations, identify potential redundant features, and further refine the feature selection.²¹ To further reduce feature dimensionality and select the most representative features, we employed the least absolute shrinkage and selection operator (LASSO) algorithm with 10-fold cross-validation.²² The 10-fold cross-validation technique was utilized to precisely tune the penalty parameter, ensuring the model’s robust generalization capability. The Lasso regression analysis was implemented using the scikit-learn package in Python, a widely recognized and powerful machine learning library. This implementation allowed for efficient computation and seamless integration with other data processing and analysis steps in our research pipeline.

Establishment and evaluation of radiomics models

Following Lasso feature selection, we input the final features into machine learning models. This study employed three machine learning models: Support Vector Machine (SVM), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM).^23,24 These models were used to construct separate diagnostic models for SLAP lesions.

Radiomics features extracted from 149 shoulder MRI scans were used for model construction. 104 imaging datasets were randomly assigned to the training set for building the SLAP lesion diagnostic models, while the remaining 45 were allocated to the test set for evaluating diagnostic performance.

The predictive performance of the models was assessed using Receiver Operating Characteristic (ROC) curves, with the AUC calculated.²⁵ We also evaluated the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of each model. Sensitivity refers to the ability to correctly identify patients with the condition as determined by the gold standard, also known as the true positive rate.²⁶ Specificity is the ability to correctly identify those without the condition, also known as the true negative rate.²⁷ The PPV represents the probability of actually having the disease when the model predicts a positive result, while the NPV represents the probability of not having the disease when the model predicts a negative result. Additionally, Decision Curve Analysis (DCA) was conducted to assess the clinical utility of each model by calculating the net benefit across various threshold probabilities, comparing model performance against the strategies of treating all or no patients.²⁸

Evaluation of traditional manual diagnosis

Two radiologists with experience in musculoskeletal imaging independently analyzed the shoulder MRI images and determined the presence or absence of SLAP lesions. Disagreements were resolved through consensus after joint review. Data analysis was performed using Python software packages to calculate the sensitivity and specificity of traditional manual diagnosis. The diagnostic performance of traditional manual diagnosis was evaluated using ROC curves, and the AUC values were used to compare the performance of traditional manual diagnosis with that of the radiomics models.

Nomogram of the combined model

This study collected key clinical information related to SLAP injuries, including basic information and medical history such as age, gender, history of shoulder joint injury and duration of symptoms, imaging results interpreted by radiologists, and whether full-thickness rotator cuff tear were observed arthroscopically. These variables represent several important aspects that surgeons routinely focus on in daily practice. Using a significance level of p < 0.05, univariate logistic regression analysis identified two statistically significant variables: age and radiologist-interpreted imaging results (Table 1). All six clinical variables were then entered into a multivariable logistic regression model (Table 2) to assess their independent contributions. Subsequently, these six clinical variables were integrated with radiomics features to construct the combined diagnostic model. In this study, we developed a combined diagnostic model by integrating predictions from the optimal radiomics model (defined by highest AUC in the test set) and clinical characteristics, visually represented as a nomogram, and evaluated its diagnostic efficacy. By integrating predictive factors from diverse clinical data, nomograms provide robust support for personalized medicine. Their application has demonstrated considerable value in predicting disease risk, evaluating treatment efficacy, and monitoring disease progression.²⁹ The diagnostic performance of the combined model was evaluated in both training and test cohorts using sensitivity, specificity, and ROC curve analysis.

Table 1.

Univariable logistic regression analysis of clinical variables associated with SLAP lesion.

Variable	OR	Lower 95%CI	Upper 95%CI	p-value
Age	1.05	1.01	1.09	0.013
Gender
Woman	1(Ref)
Man	0.82	0.42	1.6	0.567
HT
No	1(Ref)
Yes	0.97	0.51	1.87	0.938
SD	1.01	0.99	1.04	0.394
FTRCT
No	1(Ref)
Yes	1.77	0.92	3.41	0.086
MRI-BD
No	1(Ref)
Yes	2.64	1.23	5.67	0.013

Note. HT: History of trauma; SD: Symptom duration (months); FTRCT: full-thickness rotator cuff tear; MRI-BD: MRI-based diagnosis. 1(Ref) denotes the reference category used as the baseline for calculating odds ratios. For categorical variables, ORs represent the odds relative to the reference group. For continuous variables, ORs represent the change in odds per unit increase. Statistical significance was defined as p < 0.05.

Table 2.

Multivariable logistic regression analysis of clinical variables associated with SLAP lesion.

Variable	OR	Lower 95%CI	Upper 95%CI	p-value
Age	1.038	0.999	1.079	0.054
Gender	0.924	0.457	1.869	0.826
HT	0.882	0.441	1.766	0.724
SD	1.013	0.988	1.04	0.315
FTRCT	1.69	0.838	3.408	0.143
MRIBD	2.686	1.2	6.009	0.016

Note. HT: History of trauma; SD: Symptom duration (months); FTRCT: full-thickness rotator cuff tear; MRI-BD: MRI-based diagnosis. Statistical significance was defined as p < 0.05.

Result

Data results

This study retrospectively collected clinical data from patients who underwent MRI scans and shoulder arthroscopic surgery at our hospital between January 1, 2019, and December 31, 2024. A total of 268 patients were initially collected. After applying inclusion and exclusion criteria, 149 patients were ultimately included in the study. The diagnosis of SLAP lesions was based on arthroscopic findings during shoulder arthroscopic surgery. Among the included patients, 66 cases were diagnosed with SLAP lesions, while 83 cases were without SLAP lesions.

In the test set, there was a statistically significant difference in age between the groups. Other clinical indicators showed no statistical differences. In the training set, there was a statistically significant difference in the artificial diagnosis of SLAP lesions based on MRI. Other clinical indicators showed no statistical differences (Table 3).

Table 3.

Demographic and clinical characteristics of patients in training and test sets.

Name of feature	Training set (n = 104)			Test set (n = 45)
Name of feature	Uninjured	Injured	p-value	Uninjured	Injured	p-value
Age (years, mean ± SD)	59.69 ± 10.78	63.31 ± 7.26	0.121	60.67 ± 9.58	65.81 ± 10.98	0.04
Gender
man	24 (40.678)	16 (35.556)	0.742	10 (41.667)	8 (38.095)	1
woman	35 (59.322)	29 (64.444)		14 (58.333)	13 (61.905)
History of trauma, n (%)			0.948			0.824
No	26 (44.07)	21 (46.67)		11 (45.83)	8 (38.10)
Yes	33 (55.93)	24 (53.33)		13 (54.17)	13 (61.90)
Symptom duration (months, mean ± SD)	9.37 ± 13.54	9.93 ± 11.88	0.6	7.95 ± 7.78	12.77 ± 19.38	0.927
Arthroscopic full-thickness rotator cuff tear, n (%)			0.191			0.616
No	35 (59.32)	20 (44.44)		12 (50.00)	8 (38.10)
Yes	24 (40.68)	25 (55.56)		12 (50.00)	13 (61.90)
MRI-based diagnosis, n (%)			0.03			0.533
Positive	51 (86.44)	30 (66.67)		18 (75.00)	13 (61.90)
Negative	8 (13.56)	15 (33.33)		6 (25.00)	8 (38.10)

Feature statistics

From each patient, 107 radiomics features were extracted, including 18 first-order features, 14 shape features, and the remaining texture features. The methods used to extract texture features from the images included: GLCM, GLRLM, GLSZM, NGTDM, and Gray Level Dependence Matrix (GLDM). Table 4 shows the number of various features extracted. Figure 4(a) shows the percentage of various features relative to all extracted features, and Figure 4(b) presents the P-values of all features obtained from the Mann–Whitney U test or the independent samples t-test.

Table 4.

Radiomic features.

Name of feature	Number
First-order	18
GLCM	24
GLDM	14
GLRLM	16
GLSZM	16
NGTDM	5
Shape	14

Note. First-order: First Order Features; GLCM: Gray Level Co-occurrence Matrix; GLDM: Gray Level Dependence Matrix; GLRLM: Gray Level Run Length Matrix; GLSZM: Gray Level Size Zone Matrix; NGTDM: Neighboring Gray Tone Difference Matrix; Shape: Shape Features.

Figure 4.

Distribution of all radiomics features and their p values.

After the initial filtering process based on P-values, 22 features were identified. This number was subsequently reduced to 19 features based on Pearson correlation analysis. The LASSO regression procedure finally reduced the number of features to only 9, which were used for further analysis. Lasso Feature Selection Process: We conducted a 10-fold cross-validation to determine the optimal penalty coefficient λ. The Lasso cross-validation curve and regression coefficient path are illustrated in Figure 5. As evident from the graph, the optimal λ value was ascertained to be 0.0222. The features corresponding to non-zero coefficients at this λ value were retained for subsequent regression model fitting.

Figure 5.

Parameter selection and feature selection for the Lasso regression model.

The selected features encompass: Two first-order features (original_firstorder_Kurtosis and original_firstorder_Skewness), four texture features (including gray-level co-occurrence matrix features: original_glcm_ClusterShade and original_glcm_MCC, as well as gray-level size zone matrix features: original_glszm_LargeAreaHighGrayLevelEmphasis and original_glszm_SmallAreaEmphasis), and three shape features (original_shape_Flatness, original_shape_Maximum2DDiameterColumn, and original_shape_Sphericity). Figure 6 illustrates the nine selected radiomics features and their corresponding coefficients in the Lasso model.

Figure 6.

Selected radiomics features and their coefficients.

Machine learning models based on radiomics features

Table 5 presents all models for diagnosing SLAP injuries based on radiomics features. In both the training and test sets, the SVM, RF, and LightGBM models demonstrated high predictive performance with AUC values >0.800 and accuracy >0.700 (see Figure 7).

Table 5.

The diagnostic performance of the three radiomics models.

Name of the model		AUC	Accuracy	Sensitivity	Specificity	Positive predictive value	Negative predictive value
SVM	Training set	0.856	0.75	0.911	0.627	0.651	0.902
SVM	Test set	0.813	0.733	0.667	0.792	0.737	0.731
RF	Training set	0.931	0.865	0.844	0.881	0.844	0.881
RF	Test set	0.802	0.733	0.714	0.75	0.714	0.75
LightGBM	Training set	0.896	0.817	0.844	0.797	0.76	0.87
LightGBM	Test set	0.867	0.778	0.952	0.625	0.69	0.937

Figure 7.

ROC curves of the three machine learning models evaluated on the training and testing datasets.

In the test set: The SVM model showed the highest specificity among the three models at 0.792. The LightGBM model exhibited the highest sensitivity at 0.952. Comparing the diagnostic performance of the three machine learning models based on the test set AUC: The LightGBM model demonstrated the best performance, achieving the highest AUC value of 0.867 in the test cohort. Notably, the test set AUC (0.867) for the LightGBM model was slightly lower than its training set AUC (0.896), indicating good generalization without significant overfitting.

DCA on the independent test cohort (Figure 8) validated the clinical utility observed in the training cohort (Figure 9). All three models (SVM, RF, and LightGBM) demonstrated positive net benefit compared to treating all or no patients across clinically relevant threshold probabilities.

Figure 8.

DCA curves of the three machine learning-based models in the test sets.

Figure 9.

DCA curves of the three machine learning-based models in the training sets.

In the test cohort, the RF model maintained its superior performance, showing the highest and most stable net benefit across the majority of threshold probabilities ranging from 0.1 to 0.8. The LightGBM model exhibited comparable performance to RF, particularly in the moderate threshold probability range (0.2–0.6), while demonstrating slightly lower net benefit at higher thresholds. The SVM model, although showing positive net benefit, displayed more limited clinical utility with a narrower range of beneficial threshold probabilities compared to the other two models.

Diagnostic value of traditional manual diagnosis

Two radiologists with experience in musculoskeletal imaging independently analyzed the shoulder MRI images. Among 149 patients, there were 23 true positives, 52 true negatives, 14 false positives, and 60 false negatives. The accuracy of MRI plain scan for the artificial diagnosis of SLAP injury was 50.3%, with a diagnostic sensitivity of 27.7%, specificity of 78.8%, and AUC of 0.619. The traditional manual use of MRI plain scans for diagnosing SLAP injuries has limited effectiveness.

Nomogram

The integrated diagnostic model, amalgamating predictions from the lightGBM model and clinical characteristics, has been visualized in the form of a nomogram, as illustrated in Figure 10. In the training cohort, the combined model achieved an impressive AUC of 0.911. When applied to the test group, the model maintained robust performance with an AUC of 0.899. The training set model demonstrated a specificity of 83.1% and sensitivity of 84.4%, while the testing set model exhibited a specificity of 91.7% and sensitivity of 76.2%.

Figure 10.

Nomogram of the combined model.

Discussion

The mechanism of SLAP lesions is quite complex, and current research still lacks a comprehensive understanding. At present, there is a lack of effective methods for diagnosing SLAP lesions in clinical practice.^30–34 This study developed a diagnostic model for SLAP lesions based on MRI radiomics. By analyzing nine radiomic features highly correlated with SLAP lesions, the diagnostic model we constructed achieved an AUC of 0.867, sensitivity of 0.952, and specificity of 0.625 on the test set. Furthermore, when comparing the diagnostic performance of this model with traditional manual diagnostic methods, the results showed that the machine learning model outperformed manual diagnosis. Additionally, we developed a combined model incorporating both clinical and radiomic features, presented as a visual nomogram. This model exhibited strong diagnostic ability on the test set evaluation, with an AUC of 0.899, sensitivity of 0.762, and specificity of 0.917. This further confirms the excellent diagnostic capability of the radiomics model for SLAP lesions.

Imaging examinations play an irreplaceable role in the preoperative identification and evaluation of SLAP lesions. However, at present, radiologists can only provide qualitative information and semi-quantitative data from medical images, without the ability to further quantify image content. This limitation results in a vast amount of imaging data remaining underutilized.^35,36 Radiomics, by contrast, treats images as mineable quantitative data, capable of extracting information that may not be discernible to the naked eye—essentially, “images are more than pictures, they are data”.^37,38 The extracted radiomic features quantitatively describe aspects such as intensity distribution, spatial relationships, textural heterogeneity, lesion morphology, and the interactions between lesions and surrounding tissues.³⁹ These features can then be linked to clinically relevant outcomes, paving new avenues for the application of artificial intelligence in precision diagnosis. In recent years, radiomics has shown tremendous potential in the diagnosis and prognosis of orthopedic diseases.^22,40 Varriano et al.⁴¹ reported that their radiomics-based model demonstrated superior diagnostic performance for adhesive capsulitis detection, with the automated method achieving an accuracy of 0.7455 versus 0.5409 for beginner-level radiology residents.

In the study by Fei et al.,¹⁷ a machine learning model for detecting rotator cuff tear was developed and validated based on radiomic features from the supraspinatus region. Furthermore, a predictive model for rotator cuff re-tear was developed by integrating radiomic features extracted from three distinct anatomical regions: the supraspinatus, infraspinatus, and humeral head. The predictive performance of this radiomic model in forecasting postoperative re-tear was comprehensively assessed. The findings confirmed that radiomics-based models achieved strong performance in both diagnosing rotator cuff tear and predicting postoperative re-tear, thereby opening new possibilities for applying radiomics to the diagnosis of other soft tissue disorders of the shoulder.

In the present study, our radiomics model for diagnosing SLAP lesions outperformed conventional manual diagnostic methods. Both the AUC and accuracy of the model exceeded those of manual diagnosis, underscoring the advantage of radiomics in diagnostic precision. By quantitatively analyzing lesion heterogeneity, radiomics offers a more accurate means of characterizing pathological changes compared with traditional visual interpretation. Conventional manual diagnosis is constrained by the limited capacity of human visual perception to identify subtle imaging characteristics, particularly inconspicuous radiological features, which consequently reduces diagnostic precision. Radiomics, by quantifying imaging features and applying standardized evaluation criteria, enables the more sensitive and precise detection of such subtle differences.

In a study conducted by Kibler et al.,⁴² the clinical experience of surgeons in diagnosing SLAP lesions was investigated. The results showed that 57% of surveyed surgeons regarded arthroscopy as the most accurate diagnostic method, while 36% considered clinical history to be the most important factor, followed by imaging findings and physical examination. This underscores the significance of clinical history in the diagnostic process.

The results of this study demonstrated that the combined radiomics–clinical model outperformed the radiomics model. This may be explained by the fact that the MRI-based radiomics model exclusively captured pathological and anatomical information of the superior labrum, without incorporating other clinically relevant dimensions. Since the development of SLAP lesions is a dynamic process, integrating clinical indicators provides a more comprehensive reflection of the disease profile: patient age reflects the degree of labral degeneration. Radiological interpretation by attending radiologists or surgeons represents a pivotal diagnostic reference in the clinical assessment of SLAP lesions. In this investigation, radiological diagnostic determinations were incorporated as a fundamental clinical variable for model construction. The inclusion of such multidimensional information substantially enhances the predictive capacity of radiomics models. At present, there is no consensus regarding the clinical presentation and imaging standards for diagnosing SLAP lesions. Therefore, in clinical practice, diagnostic decisions should be made by synthesizing multiple sources of evidence rather than relying on a single modality.

This study has several limitations that should be acknowledged. First, the retrospective single-center design and relatively small, unevenly distributed sample size may have introduced bias and increased the risk of model overfitting. To enhance generalizability and stability, future work should involve larger cohorts and multi-center external validation. Second, image segmentation in this study relied on manual delineation of regions of interest by experienced radiologists. Although this ensured expert anatomical input, manual segmentation is inherently time-consuming and operator dependent, which limits practicality for real-world use and may reduce reproducibility. Moreover, the present study does not provide an automated or clinically feasible workflow for implementation. Third, the clinical features included in this study did not encompass specialized physical examination maneuvers specifically designed for the assessment of SLAP lesions. The absence of these diagnostic indicators may have limited the predictive performance of the clinical model. Incorporating targeted physical examination tests in future studies could strengthen diagnostic accuracy and enhance clinical applicability. Although our models demonstrated improved diagnostic performance compared with routine interpretation of non-contrast MRI by radiologists, clinical management of suspected SLAP lesions is multifactorial and should not rely on imaging findings alone. In contemporary practice, treatment decisions are primarily guided by patient age, symptom profile, activity demands, concomitant shoulder pathology, and response to nonoperative management. Finally, although DCA suggests clinical utility across relevant risk thresholds, this study is designed solely as an adjunctive decision-support tool to reduce diagnostic uncertainty and does not demonstrate how improved diagnostic accuracy would alter surgical indications or reduce unnecessary arthroscopies. Subsequent investigations should incorporate validated patient reported outcome measures specific to the shoulder to determine whether pathways guided by the model achieve clinically meaningful improvements, interpreted using established minimal clinically important difference (MCID) or Patient Acceptable Symptom State (PASS) thresholds.

Conclusion

In summary, this study developed and validated machine learning models based on both clinical characteristics and radiomics features for the diagnosis of SLAP lesions. The radiomics model, constructed from shoulder MRI-derived features, demonstrated effective identification of SLAP lesions and achieved higher diagnostic accuracy compared with conventional manual assessment. Moreover, the combined model integrating radiomics and clinical features further enhanced predictive performance for SLAP lesion risk and was visualized through a nomogram, offering a practical tool for clinical application.

Footnotes

ORCID iD

Hongyu Wang

Ethical considerations

This retrospective study was approved by the hospital ethics committee. Due to the retrospective nature of the study using anonymized data,informed consent was waived by the ethics committee.

Funding

The authors received no financial support for the research,authorship,and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

References

Snyder

Banas

Karzel

. An analysis of 140 injuries to the superior glenoid labrum. J Shoulder Elb Surg 1995; 4: 243–248.

Onyekwelu

Khatib

Zuckerman

, et al. The rising incidence of arthroscopic superior labrum anterior and posterior (SLAP) repairs. J Shoulder Elb Surg 2012; 21: 728–731.

Kütük

Kanatli

Ataoğlu

, et al. The correlation of the SLAP II lesion findings between physical examination, magnetic resonance imaging, and arthroscopic surgery. Indian J Orthop 2021; 55: 443–448.

Dean

Onsen

Lima

, et al. Physical examination maneuvers for SLAP lesions: a systematic review and meta-analysis of individual and combinations of maneuvers. Am J Sports Med 2023; 51: 3042–3052.

Gismervik

Drogset

Granviken

, et al. Physical examination tests of the shoulder: a systematic review and meta-analysis of diagnostic test performance. BMC Muscoskelet Disord 2017; 18: 41.

Cook

Beaty

Kissenberth

, et al. Diagnostic accuracy of five orthopedic clinical tests for diagnosis of superior labrum anterior posterior (SLAP) lesions. J Shoulder Elb Surg 2012; 21: 13–22.

Hegedus

Goode

Cook

, et al. Which physical examination tests provide clinicians with the most value when examining the shoulder? Update of a systematic review with meta-analysis of individual tests. Br J Sports Med 2012; 46: 964–978.

Meserve

Cleland

Boucher

. A meta-analysis examining clinical test utility for assessing superior labral anterior posterior lesions. Am J Sports Med 2009; 37: 2252–2258.

Calvert

Chambers

Regan

, et al. Special physical examination tests for superior labrum anterior posterior shoulder tears are clinically limited and invalid: a diagnostic systematic review. J Clin Epidemiol 2009; 62: 558–563.

10.

Symanski

Subhas

Babb

, et al. Diagnosis of superior labrum anterior-to-posterior tears by using MR imaging and MR arthrography: a systematic review and meta-analysis. Radiology 2017; 285: 101–113.

11.

Connolly

Schwartzberg

Reuss

, et al. Sensitivity and specificity of noncontrast magnetic resonance imaging reports in the diagnosis of type-II superior labral anterior-posterior lesions in the community setting. J Bone Joint Surg Am 2013; 95: 308–313.

12.

Tuite

Rutkowski

Enright

, et al. Width of high signal and extension posterior to biceps tendon as signs of superior labrum anterior to posterior tears on MRI and MR arthrography. AJR Am J Roentgenol 2005; 185: 1422–1428.

13.

Arirachakaran

Boonard

Chaijenkij

, et al. A systematic review and meta-analysis of diagnostic test of MRA versus MRI for detection superior labrum anterior to posterior lesions type II-VII. Skelet Radiol 2017; 46: 149–160.

14.

Wang

Wen

Yin

, et al. Musculoskeletal ultrasound image-based radiomics for the diagnosis of Achilles tendinopathy in skiers. J Ultrasound Med 2023; 42: 363–371.

15.

Chen

Wang

Zhu

, et al. A novel application of unsupervised machine learning and supervised machine learning-derived radiomics in anterior cruciate ligament rupture. Risk Manag Healthc Policy 2021; 14: 2657–2664.

16.

Oeding

Pareek

Nieboer

, et al. A machine learning model demonstrates excellent performance in predicting subscapularis tears based on pre-operative imaging parameters alone. Arthroscopy 2024; 40: 1044–1055.

17.

Fei

Wan

, et al. Novel methods to diagnose rotator cuff tear and predict post-operative re-tear: radiomics models. Asia Pac J Sports Med Arthrosc Rehabil Technol 2024; 37: 14–20.

18.

Yushkevich

Piven

Hazlett

, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 2006; 31: 1116–1128.

19.

Lam

SWC

. Texture feature extraction using gray level gradient based co-occurence matrices. 1996 IEEE Int Conf Syst Man Cybern Inf Intell Syst 1996; 261: 267–271.

20.

van Griethuysen

JJM

Fedorov

Parmar

, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res 2017; 77: e104–e107.

21.

Wang

Peng

Feng

, et al. Development and validation of a computed tomography-based radiomics signature to predict response to neoadjuvant chemotherapy for locally advanced gastric cancer. JAMA Netw Open 2021; 4: e2121143.

22.

Cui

Liu

Jing

, et al. Development of machine learning models aiming at knee osteoarthritis diagnosing: an MRI radiomics analysis. J Orthop Surg Res 2023; 18: 375.

23.

Chen

Yao

, et al. Application of machine learning algorithms to identify people with low bone density. Front Public Health 2024; 12: 1347219.

24.

Chen

Pan

, et al. US-based radiomics analysis of different machine learning models for differentiating benign and malignant BI-RADS 4A breast lesions. Acad Radiol 2025; 32: 67–78.

25.

Ouyang

, et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: a machine learning, multicenter study. EBioMedicine 2021; 69: 103460.

26.

Jajroudi

Enferadi

Homayoun

, et al. MRI-based machine learning for determining quantitative and qualitative characteristics affecting the survival of glioblastoma multiforme. Magn Reson Imaging 2022; 85: 222–227.

27.

Zhang

Zhong

Liang

, et al. MRI-based radiomic and machine learning for prediction of lymphovascular invasion status in breast cancer. BMC Med Imag 2024; 24: 322.

28.

Zhou

, et al. Development of a radiomics and clinical feature-based nomogram for preoperative prediction of pathological grade in bladder cancer. Front Oncol 2025; 15: 1661979.

29.

Balachandran

Gonen

Smith

, et al. Nomograms in oncology: more than meets the eye. Lancet Oncol 2015; 16: e173–e180.

30.

Bryan

White

Inclan

, et al. A comprehensive review of the physical examination for the biceps-labrum complex of the shoulder. Arthrosc Tech 2024; 13: 103136.

31.

Yang

Kim

, et al. Understanding the physical examination of the shoulder: a narrative review. Ann Palliat Med 2021; 10: 2293–2303.

32.

Knesek

Skendzel

Dines

, et al. Diagnosis and management of superior labral anterior posterior tears in throwing athletes. Am J Sports Med 2013; 41: 444–460.

33.

Amin

Youssef

. The diagnostic value of magnetic resonance arthrography of the shoulder in detection and grading of SLAP lesions: comparison with arthroscopic findings. Eur J Radiol 2012; 81: 2343–2347.

34.

Sheridan

Kreulen

Kim

, et al. Accuracy of magnetic resonance imaging to diagnose superior labrum anterior-posterior tears. Knee Surg Sports Traumatol Arthrosc 2015; 23: 2645–2650.

35.

Yip

Aerts

. Applications and limitations of radiomics. Phys Med Biol 2016; 61: R150–R166.

36.

Nougaret

Tibermacine

Tardieu

, et al. Radiomics: an introductory guide to what it may foretell. Curr Oncol Rep 2019; 21: 70.

37.

Gillies

Kinahan

Hricak

. Radiomics: images are more than pictures, they are data. Radiology 2016; 278: 563–577.

38.

Rogers

Thulasi Seetha

Refaee

TAG

, et al. Radiomics: from qualitative to quantitative imaging. Br J Radiol 2020; 93: 20190948.

39.

Ibrahim

Primakov

Beuque

, et al. Radiomics for precision medicine: current challenges, future prospects, and the proposal of a new framework. Methods 2021; 188: 20–29.

40.

Cao

Tong

, et al. Artificial intelligence in thyroid ultrasound. Front Oncol 2023; 13: 1060702.

41.

Varriano

Nardone

Brunese

, et al. An approach leveraging radiomics and model checking for the automatic early diagnosis of adhesive capsulitis. Sci Rep 2024; 14: 18878.

42.

Kibler

Sciascia

. Current practice for the diagnosis of a SLAP lesion: systematic review and physician survey. Arthroscopy 2015; 31: 2456–2469.

A diagnostic model based on clinical indicators and radiomics features for superior labral anterior and posterior lesions in the shoulder joint

Abstract

Background

Methods

Results

Conclusion

Keywords

Introduction

Method

Methods and tools

Participants in the study

Collection of patient clinical information

Research methodology

Magnetic resonance image collection and delineation of target regions

Extraction and screening of radiomics features

Establishment and evaluation of radiomics models

Evaluation of traditional manual diagnosis

Nomogram of the combined model

Result

Data results

Feature statistics

Machine learning models based on radiomics features

Diagnostic value of traditional manual diagnosis

Nomogram

Discussion

Conclusion

Footnotes

ORCID iD

Ethical considerations

Funding

Declaration of conflicting interests

References