MRI-based Radiomic Analysis Of Soft Tissue Reactions Near Total Hip Arthroplasty

Abstract

This study applied radiomics to MRI data for automated classification of soft tissue abnormalities near total hip arthroplasty (THA). A total of 126 subjects with 1.5T MRI of symptomatic THA were included in the analysis. Peri-prosthetic soft tissue regions of interest were manually segmented and classified by an expert radiologist. An established radiomics library was used to extract 96 features from 2D image patches across segmented regions. Logistic regression was employed as the primary radiomic classifier, achieving an average area under curve (AUC) of 0.71 in differentiating tissue classifications spanning normal, infected, and several inflammatory, noninfectious categories. Notably, infection cases were identified with the highest accuracy, attaining an AUC of 0.79. Statement of Clinical Significance: This study demonstrates that radiomics applied to MRI data can effectively automate the classification of soft tissue abnormalities in symptomatic total hip arthroplasty, particularly in differentiating periprosthetic infections.

INTRODUCTION

Total hip arthroplasty (THA) is an effective treatment for alleviating pain and restoring function in patients suffering from end-stage degenerative joint disease. In the United States alone, the demand for primary THA procedures s projected to increase by 17% to 572,000 by 2030 [1]. Revision procedures are also increasing steadily due to short and long-term complications after primary procedures, with projection models estimating that revision hip arthroplasty (RHA) procedures will exceed 100,000 annually in the United States by 2030 [1]. The most common indications for THA revision include periprosthetic osteolysis, aseptic mechanical loosening, instability, infection, femoral component loosening, and adverse local tissue reactions [2].

Periprosthetic tissue damage and inflammation are frequently observed in patients undergoing THA revision. Magnetic resonance imaging (MRI) has emerged as the advanced imaging modality of choice for comprehensively assessing soft tissue conditions around total hip arthroplasty (THA), owing to its superior soft tissue contrast resolution, lack of ionizing radiation, and multi-planar capabilities. Historically, MRI of THA has been challenging due to metallic susceptibility artifact in imaged implant regions. These artifacts manifest as voxel pile-up, signal voids and image warping in image reconstructions [3]. Multi-spectral pulse sequences [4–6] were developed to produce diagnostic images in the presence of metal hardware [7], providing detailed characterization of the periprosthetic environment in clinically viable scan times [8].

MRI facilitates clear visualization and identification of periprosthetic bone loss and adverse host-mediated inflammatory reactions in soft tissues [9–14]. Multiple prior studies have surgically validated that relevant morphological [15] and quantitative MRI signatures [16] correlate well with underlying tissue pathology. Distinct from blood serum levels, MRI can directly visualize inflammatory reactions and determine the extent and degree of attendant soft-tissue damage. However, the utility of blood serum measures is limited, as patients can demonstrate elevated serum metal ion levels without any observed inflammatory soft-tissue reactions on MRI [17]. Furthermore, patients with low serum ion levels may show substantial soft-tissue inflammatory responses on MRI [15].

Accurate and consistent evaluation of THA using MRI requires subspecialty expertise, making radiological classification of synovial abnormalities challenging. A review by Fritz et al. [18] effectively displays this complexity with a 30-element descriptive chart characterizing the image signatures of osteolysis and synovitis near THA, including the nuanced impact and interpretations of joint fluid spaces, synovial-wall thicknesses, and soft-tissue characteristics. In addition to the inherent challenges of imaging near arthroplasty and unique complications commonly encountered, the overall visual presentation and diagnostic characteristics of MRI near THA are unlike other radiologic orthopedic assessments. Precise differentiation of adverse soft tissue reactions occurring in periprosthetic tissues is critical for diagnosing and managing failing THA. This underscores an unmet need for automated quantitative technologies to perform these tasks effectively.

In recent years there has been growing interest in extracting quantitative imaging biomarkers from standard-of-care medical images through an analysis approach known as “radiomics””. Radiomics involves the high-throughput extraction and analysis of a large numbers of engineered features from medical images using data characterization algorithms [19]. These features quantify phenotypic characteristics such as shape, intensity distribution and texture patterns within the images. Radiomics converts conventional images into quantitative mineable data and can enable decision support across diverse use cases in oncology, including diagnosis, prognosis and prediction of response to treatment [20, 21]. Though potential applications of radiomics in musculoskeletal imaging has been under-explored, recent work has seen this approach deployed to characterize knee osteoarthritis [22, 23] and assess vertebral texture changes as a marker for low back pain on MRI [24].

As a preliminary exploration of applying radiomic metrics for orthopedic imaging, the present study utilized radiomic feature generation and modeling to characterize the synovial response in patients with THA. Radiomic features derived from conventional multispectral MRI metal-artifact suppressed scans were used to model periprosthetic soft tissue reactions near THA using radiologist-derived training classifications. We hypothesized that MRI phenotypes, derived from radiomic analysis, can accurately categorize expert-derived classification of synovial tissue pathology following THA.

Precise differentiation of adverse soft tissue reactions displayed in periprosthetic tissues near total hip arthroplasty (THA) could improve management of THA. By employing MRI radiomics to classify soft tissue reactions such as infection, metallosis, polymeric debris, and ALTR near total hip replacements, clinicians can obtain quantitative data on the tissue characteristics and the spatial distribution of these reactions. Radiomic characterization of these reactions could lead to more consistent THA revision planning by providing detailed imaging-based tissue characterization to enable personalized pre-operative planning and component selection. Beyond diagnosis, radiomic soft tissue characterization also offers prognostic and monitoring advantages over subjective evaluation alone.

METHODS

Study Design:

This study analyzed MRI datasets collected within a prospective research study of symptomatic THA (n=103) [15] as well as datasets collected within a retrospective study of clinical THA evaluations (n=23) at a second clinical research site.

Imaging data was collected via separately approved Institutional Review Board (IRB) protocols at two academic medical centers. Subjects within the prospective data collection arm of the dataset at site 1 provided written consent to participate in the study while the retrospective data collection arm of the study at site 2 was granted a waiver of consent. The cumulative retrospective analysis of these cohorts for the present study is classified at Evidence Level III.

The analysis and reporting of this study adhere to the guidelines outlined in the CheckList for EvaluAtion of Radiomics research (CLEAR) checklist, a recently developed documentation standard for radiomic research focusing on repeatability, reproducibility, and transparency [25].

Subjects:

A total of 126 participants who had undergone THA were analyzed (74 females, 52 males) with an average age of 67.86±9.20 years, specifically 69.04±8.23 years for males and 67.03±9.74 years for females. All subjects were experiencing symptoms (reported pain) related to their THA at the time of data collection. The time between imaging exam and implantation across the cohort was 9.0 +/− 7.1 years.

For both data collection arms of these study, the inclusion criteria were adults (aged 18 years or older) who had undergone total hip replacement surgery and were experiencing postoperative symptoms. No available data were specifically removed from analysis due to any exclusion criteria.

MRI Acquisition and Analysis:

All evaluated images were acquired using 1.5T MRI systems: Optima 450 and Optima 450w (GE Healthcare). Coronal acquisitions of proton density (PD) and short-tau inversion recovery (STIR) contrast images using 3DMSI (MAVRIC SL, GE Healthcare) metal artifact suppression technology were leveraged for analysis. The STIR and PD images were selected for the analysis since they provide complementary information, with STIR offering superior fluid sensitivity but lower resolution as compared to PD. The acquisition parameters for these scans followed established standard of care recommendations [26] and are summarized in Table 1.

Table 1:

Key acquisition parameters for 3D-MSI PD and STIR images utilized for radiomic analysis.

Parameter	PD	STIR
Scan Plane	Coronal	Coronal
Localization	Whole pelvis	Whole pelvis
3D-MSI Version	MAVRIC SL	MAVRIC SL
Repetition time (msec)	3–5 s	4–5 s
Echo time (msec)	8	8
Inversion time (ms)	None	150
Echo train length	20–24	20–24
Receiver bandwidth (Hz per pixel)	488.3	976.6
FOV (cm)	40 x 40	40 x 40
Acquired In-Plane Matrix	512 x 256	256 x 192
Slice thickness (mm)	4–5 mm	5
Number of phase-encoded slices	24–36	24–36
Acquisition time (min)	4–8 min	4–8 min

Radiologist-based region-of-interest segmentations and classifications were performed by an expert sub-specialized orthopaedic radiologist with 20 years of experience in imaging near arthroplasty. Three-dimensional segmented regions of interest based on the extent of the synovial region were traced on a slice-wise basis. The size of these regions was variable across the participants but was not explicitly dependent on implant size. Synovial classifications were categorized into 6 categories using specific radiological criteria. This classification scheme has been previous reported [15], but is repeated here for contextual purposes:

• Normal: a thin capsule with low signal intensity;

• ALTR (adverse local tissue reaction): a thickened, hyperintense capsule, often lacking a clear boundary with the surrounding muscle signal and tissue structure, indicating necrosis;

• Metallosis: low signal intensity deposits found in the capsular lining inside the joint or in an extracapsular location;

• Infection: layered synovial lining with surrounding edema;

• Polymeric: intracapsular concentrations of particulate, intermediate signal intensity debris; and

• Abnormal: maximum inferomedial synovial thickness in the coronal plane.

Grading was carried out in fully a blinded manner, without knowledge of the implant design and composition, corresponding radiographs, reason for examination (retrospective cohort), or reason for revision, subsequent histologic, implant wear, or implant corrosion analysis (prospective cohort). The inter- and intra-rater agreement of MR synovial classification was previously assessed between two independent readers and found an inter-rater agreement of substantial to almost perfect (Gwet’s AC1 range, 0.65–0.97) and an intra-rater agreement of moderate to almost perfect (Gwet’s AC1 range, 0.59–0.99) [15].

Image Registration and Resampling:

The acquired STIR and PD image volumes were aligned to enable combined radiomic analysis across both imaging contrasts. Registration was performed using the antsRegistration module in ANTsPy [27], an open-source toolkit for image registration and segmentation. The utilized registration approach implements the symmetric image normalization method (SyN) algorithm for deformable registration. The PD volumes were resampled to match the lower resolution of the STIR volumes using third order B-spline interpolation as implemented in NiBabel [28].

Overlapping 2D patches of three sizes: [12 x 12, 20 x 20, 30 x 30] pixels corresponding to patch square patch areas of [0.9, 2.4, 5.5] cm² were extracted across each segmented synovial reaction from the registered and resampled STIR and PD volumes. Patches were constructed to overlap by 50% to provide dense coverage and capture boundary information.

Radiomic features were extracted from the 2D patches using the open-source PyRadiomics (v3.0) [29] package in Python. A total of 96 first-order and texture-based radiomic features were calculated, including statistical metrics such as mean, median, skewness, and kurtosis. Texture features, including contrast, correlation, energy, and homogeneity were derived from the Gray Level Co-occurrence Matrix (GLCM). The original images were utilized for radiomic features generation (i.e,.. no filters were applied).

Classification Modeling:

Supervised classification modeling was implemented using the scikit-learn machine learning library in Python [30]. The primary classifier was L2-regularized logistic regression with stratified 5-fold cross-validation, implemented via LogisticRegressionCV() in scikit-learn. Separate binary classification models were constructed for each of the 6 tissue reaction classes and each of the patch size data sets, setting the patches for that label as positives and the remaining patches as negatives. This enabled training models to discriminate each tissue type against all others. The current dataset was insufficient for training numerically stable multi-class classification models.

The modeling pipeline was executed separately using the STIR or PD radiomic features alone (single contrast), and with both feature sets combined (combined contrast). The full feature set for each modeling experiment was first scaled using by removing the feature means and scaling to unit variance. Principal component analysis (PCA) was then applied to reduce the dimensionality of the feature space.

Hyperparameters, including, regularization strength were tuned during model training using inner 5-fold cross-validation. Final model performance was evaluated using nested 10-fold cross-validation, ensuring strict separation between training and testing data. The utilized model performance metric was the receiver-operator curve, summarized via the mean area under the curve (AUC) across folds. Additional metrics including accuracy, precision-recall AUC, and Cohen’s kappa were considered but did not provide significant differentiation from the ROC AUC. All metrics were computed using utilities from scikit-learn. Figure 1 provides a flow-chart for the full image analysis and modeling approaches utilized within the study.

Figure 1:

Flowchart outline procedure from acquired imaging data to performance metrics of classification models

Results

Figure 2 displays examples of radiologist-based segmentations of synovial regions categorized into the six tissue reaction classes analyzed in this study. For each case, the STIR image, PD image, and a fused overlay are shown, along with zoomed patches demonstrating the extracted texture phenotypes. The patches reveal subtle characteristic textures for each tissue type that forms the basis for radiomic feature extraction and classification.

Figure 2:

Sample images, soft tissue segmentations, and extracted image patches used for radiomic analysis within each of the studied pathology classes. Zoomed regions indicated areas of focus near segmented soft tissue areas of interest. Segmentations utilized for radiomic analyses are indicated in the far-right column for each case, where the segmented region is highlighted in semi-transparent red hue.

A total of 13454, 8548, and 5148 2D patches, with respective patch sizes of 0.9, 2.4, and 5.5 cm² . were extracted from 142 manually segmented zones across 126 subjects’ MRI scans. Patch classification distributions for each patch size setting are presented in Table 2. Normal tissue patches comprise the largest group, as expected given the underlying class prevalence. Among pathological reactions, the classification of infection offered a substantial population of patches due to the anatomic extent of the synovial reactions in this category. The variation of patch counts across the classification categories emphasizes the need for class balanced approaches when training and evaluating classifier models.

Table 2.

of radiomic classification patch counts and independent tissue evaluation zones (contiguous regions of evaluated tissue) used to construct the patched dataset

	Number of Patches			Distinct Evaluation Zones
Patch Size	0.9 cm2	2.4 cm2	5.5 cm2
Class
Normal	3734	2580	1678	91
ALTR	620	505	214	3
Infection	2933	1808	933	7
Metallosis	1460	931	606	16
Abnormal	846	559	323	14
Polymeric	3861	2165	1394	11
TOTAL	13454	8548	5148	142

PCA transformed the 96 single contrast or 192 combined contrast radiomic features down to the minimum number of components that retained over 95% of the full feature set’s explained variance. The number of features was reduced to a range of 11 to 16 features using this approach on the different model constructions. This substantial dimensionality reduction implies a high degree of redundancy in the radiomic feature sets.

Classifier performance for differentiating each of the six tissue types, assessed using cross-validated area under the receiver-operator curve (AUC), is shown in Table 3. Classification utilizing combined STIR and PD radiomic features demonstrated an average AUC of 0.65, 0.69, and 0.70 across all tissue types for respective patch sizes of 0.9, 2.4, and 5.5 cm². Performance reached 0.81 AUC in detecting the classification of infection when using combined features at the largest patch size of 5.5 cm². STIR alone also offered strong discrimination for the classification of infection to fluid signal. When using an optimal contrast for each category, the mean performance across the tissue classes was slightly increased by 0.01 for each respective patch size (maximized at 0.71 for a patch size of 5.5 cm²).

Table 3.

Modeling performance across included imaging contrasts for each synovial classification. Cross-validation AUC scores are represented as means and 95% confidence intervals across the 10 validation folds. The number of PCA components (N_PCA used for each model (providing 95% explained variance) is also reported.

		MRI Contrast
Class	Patch Size (cm2)	STIR	Proton Density	STIR + Proton Density	Max
		NPCA0.9=11	NPCA0.9=11	NPCA0.9=15
		NPCA2.4=12	NPCA2.4=12	NPCA2.4=15
		NPCA5.5=12	NPCA5.5=12	NPCA5.5=16
Normal	0.9	0.70 [0.60,0.79]	0.63 [0.54,0.71]	0.70 [0.60,0.79]	0.70
	2.4	0.74 [0.64,0.83]	0.66 [0.57,0.74]	0.74 [0.64,0.83]	0.74
	5.5	0.77 [0.70,0.85]	0.67 [0.60,0.74]	0.77 [0.70,0.84]	0.77
ALTR	0.9	0.63 [0.56,0.70]	0.45 [0.38,0.51]	0.61 [0.54,0.68]	0.63
	2.4	0.67 [0.58,0.76]	0.50 [0.44,0.56]	0.67 [0.58,0.76]	0.67
	5.5	0.69 [0.58,0.80]	0.51 [0.42,0.60]	0.70 [0.60,0.80]	0.70
Infection	0.9	0.73 [0.65,0.81]	0.66 [0.59,0.73]	0.73 [0.65,0.80]	0.73
	2.4	0.79 [0.69,0.89]	0.70 [0.63,0.78]	0.79 [0.69,0.89]	0.79
	5.5	0.81 [0.71,0.91]	0.72 [0.65,0.80]	0.81 [0.71,0.91]	0.81
Metallosis	0.9	0.60 [0.51,0.69]	0.62 [0.54,0.71]	0.60 [0.51,0.69]	0.62
	2.4	0.62 [0.52,0.72]	0.63 [0.54,0.72]	0.60 [0.50,0.70]	0.63
	5.5	0.60 [0.49,0.72]	0.62 [0.52,0.72]	0.58 [0.46,0.70]	0.62
Abnormal	0.9	0.62 [0.57,0.67]	0.61 [0.58,0.65]	0.62 [0.57,0.66]	0.62
	2.4	0.64 [0.56,0.73]	0.61 [0.54,0.68]	0.64 [0.55,0.73]	0.64
	5.5	0.65 [0.55,0.75]	0.63 [0.56,0.70]	0.65 [0.55,0.74]	0.65
Polymeric	0.9	0.66 [0.61,0.72]	0.62 [0.57,0.67]	0.65 [0.60,0.71]	0.66
	2.4	0.70 [0.64,0.75]	0.65 [0.60,0.70]	0.69 [0.64,0.74]	0.70
	5.5	0.71 [0.66,0.76]	0.67 [0.62,0.73]	0.71 [0.66,0.76]	0.71
Mean	0.9	0.66	0.60	0.65	0.66
	2.4	0.69	0.63	0.69	0.70
	5.5	0.71	0.64	0.70	0.71

For each specific reaction type, a classifier with the normal classified cases removed from the dataset was constructed to discriminate that pathology within the abnormal tissue classes (Table 4). The maximum classifier performance across the pathology classes was higher than the equivalent performance in the global modeling effort (Table 3), the most notable increases occurring in the classifications of ALTR, Metallosis, and Abnormal classes. The mean optimal classifier performance for the 5 pathology classes increased in AUC from 0.66, 0.69, and 0.70 in the global modeling approach to 0.69, 0.73, 0.75 in the pathology-only modeling approach for respective patch sizes of 0.9, 2.4, and 5.5 cm². Random Forrest classifiers were also explored and showed similar results to the utilized logistic regression approach for all evaluated classification tasks. Only the logistic regression results are presented here for preliminary demonstration purposes.

Table 4.

Modeling performance across included imaging contrasts as discriminators of specific pathology categories. Cross-validation AUC scores are represented as means and 95% confidence intervals across the 10 validation folds. The number of PCA components (N_PCA used for each model (providing 95% explained variance) is also reported.

		MRI Contrast
Class	Patch Size (cm2)	STIR	Proton Density	STIR + Proton Density	Max
		NPCA0.9=11	NPCA0.9=11	NPCA0.9=16
		NPCA2.4=12	NPCA2.4=11	NPCA2.4=16
		NPCA5.5=12	NPCA5.5=12	NPCA5.5=16
ALTR	0.9	0.55 [0.44,0.66]	0.63 [0.54,0.71]	0.70 [0.60,0.79]	0.70
	2.4	0.63 [0.55,0.72]	0.66 [0.57,0.74]	0.74 [0.64,0.83]	0.74
	5.5	0.63 [0.53,0.74]	0.67 [0.60,0.74]	0.77 [0.70,0.84]	0.77
Infection	0.9	0.72 [0.64,0.80]	0.45 [0.38,0.51]	0.61 [0.54,0.68]	0.72
	2.4	0.77 [0.66,0.87]	0.50 [0.44,0.56]	0.67 [0.58,0.76]	0.77
	5.5	0.79 [0.68,0.90]	0.51 [0.42,0.60]	0.70 [0.60,0.80]	0.79
Metallosis	0.9	0.61 [0.51,0.71]	0.66 [0.59,0.73]	0.73 [0.65,0.80]	0.73
	2.4	0.64 [0.53,0.75]	0.70 [0.63,0.78]	0.79 [0.69,0.89]	0.79
	5.5	0.63 [0.50,0.76]	0.72 [0.65,0.80]	0.81 [0.71,0.91]	0.81
Abnormal	0.9	0.64 [0.60,0.69]	0.62 [0.54,0.71]	0.60 [0.51,0.69]	0.64
	2.4	0.66 [0.58,0.75]	0.63 [0.54,0.72]	0.60 [0.50,0.70]	0.66
	5.5	0.68 [0.58,0.79]	0.62 [0.52,0.72]	0.58 [0.46,0.70]	0.68
Polymeric	0.9	0.65 [0.56,0.74]	0.61 [0.58,0.65]	0.62 [0.57,0.66]	0.65
	2.4	0.69 [0.60,0.77]	0.61 [0.54,0.68]	0.64 [0.55,0.73]	0.69
	5.5	0.70 [0.62,0.78]	0.63 [0.56,0.70]	0.65 [0.55,0.74]	0.70
Mean	0.9	0.64	0.59	0.65	0.69
	2.4	0.68	0.62	0.69	0.73
	5.5	0.69	0.63	0.7	0.75

Discussion

The study highlights the potential utility of radiomic features, derived from conventional MRI, as a non-invasive mechanism for classifying periprosthetic synovial reactions. The imaging data utilized for this classification study were collected using routine clinical imaging protocols on individuals with symptomatic THA using established and widely-available metal-artifact suppressed multi-spectral MRI methods. Using radiomic feature extraction methods, quantitative classification can be performed using these conventional images, negating the use of specialized quantitative image acquisition techniques.

Our supervised machine learning models were able to differentiate six tissue types with cross-validated performance scores that ranged from 0.62 for Metallosis and generalized Abnormal using 0.9 cm ² patch sizes to 0.81 for infected tissue using 5.5 cm² patches in the full-cohort classification. The full cohort classification included substantial numbers of normative tissue cases, enabling an abnormality binary classification test that yielded an AUC of 0.77 using only the STIR contrast images with 5.5 cm² patches. Models developed only on pathological tissues allowed for improved discrimination between the abnormal classes, with a mean AUC performance increase of 0.03, 0.04, and 0.05 for patch sizes of 0.9, 2.4, and 5.5 cm² relative to the full (normal-inclusive) analysis cohort. These findings suggest that a two-stage classification scheme may be beneficial, where tissues are first classified as Normal/Abnormal, followed by a specific pathology classifier once an ”abnormal” classification is determined.

The impact of patch-size on modeling performance was substantial. A general trend in all cases revealed improved classification capabilities using larger patch sizes. This result suggests that the spatial extent of radiomic features that distinguish the synovial reaction patterns is relatively large. Using a patch size of 5.5 cm² demonstrated strong classification performance for normal tissue from pathological tissues (0.77,Table 3) and ALTR (0.77), Infection (0.79), Metallosis (0.81), and Polymeric (0.70) tissue from other abnormal classifications (Table 4). Though the model-based tissue distinction capabilities improve with larger patches, the diagnostic utility of such a tool will diminish with increased patch sizes, as the ability to localize classified tissue reactions is inversely proportional to this parameter. The tradeoff of these considerations and impact on diagnostic capabilities will require further studies on larger clinical cohorts.

The short-tau inversion recovery (STIR) MRI contrast generally exhibited superior discrimination compared to PD, especially for the classification of Infection. We attribute this finding to sensitivity of STIR imaging for inflammation and fluid. However, the PD images performed better for Metallosis classification. The most challenging classification was the general Abnormal category, which yielded a maximum performance of 0.68 (using the STIR contrast) across the different patch sizes and full and pathology-only modeling efforts. It is noted that residual misregistration between the STIR and PD contrast image sets could negatively impact the performance of joint-contrast models. However, the similar performance observed when using both contrasts for classification, compared to the optimal single contrast, suggests this potential alignment issue was not a substantial confound.

Successful classification of synovial reactions could ultimately increase consistency and precision in THA diagnostics, enabling differentiation between similar-appearing pathologies to guide patient management. For example, discriminating infection from aseptic loosening is critical, as infected implants require surgical debridement and antibiotics administration, while aseptic loosening often result in revision arthroplasty. Radiomic characterization of these reactions could improve THA revisions, which are far less effective and more expensive than primary arthroplasty procedures [31, 32]. Patients undergoing THA revision with different pathologic presentations often require customized procedures and devices [33]. By providing detailed imaging-based tissue characterization, radiomics could enable personalized pre-operative planning and component selection. Beyond diagnosis, radiomic soft tissue classification also offers prognostic and monitoring advantages. Enhanced diagnostic and prognostic capabilities could also reduce costs by preventing unnecessary procedures. For instance, accurately classifying early soft tissue complications could prompt targeted intervention, avoiding complex revisions required for advanced disease, with a higher attendant morbidity.

While the use of a single expert radiologist for manual segmentation and classification introduced an element of subjectivity the present study, in a previous study utilizing the same synovial classification system demonstrated high levels of interrater agreement between subspecialized radiologists with at least 10 years of experience in assessing soft tissue reactions near THR [15]. In addition, in another previous study, we determined inter-examiner synovial segmentation to have an intraclass correlation coefficient of 0.99, [34]. These previous demonstrations of high inter-rater consistency in segmenting and classifying synovial regions justifies the use of a single radiologist in this preliminary feasibility study.

Traditional statistical power analysis techniques often employed for hypothesis testing may not be directly applicable to high-dimensional radiomic models. The complex nature of these models, involving multiple, often correlated features, undermines the utility of a simplistic power analysis based on individual predictors. Moreover, the predictive efficacy of the model is generally determined through composite features rather than isolated variables, further complicating the interpretation of individual coefficient power values. The primary objective of the current work was predictive accuracy, as quantified by the AUC metric. To compute these metrics in a robust fashion and avoid over-fitting, a cross-validation framework was implemented to validate model performance.

A dataset of labeled peri-prosthetic tissue regions from 126 subjects spanning two academic medical practices was sufficient to demonstrate proof-of-concept classification capabilities. To enhance model performance and generalizability, a larger and more diverse dataset spanning several clinical practices should be evaluated. Further, the incorporation of more sophisticated classification schemes that utilize deep convolutional neural networks could considerably improve on the promising preliminary MRI radiomic results, as well as expedite the segmentation process.

Further research could also explore correlations between radiomic signatures wand histopathologic findings, assess radiomic stability across scanners, and explore prognostic capabilities of radiomic models.

To summarize, this study demonstrated the potential of MRI radiomics to classify periprosthetic soft tissue reactions after total hip arthroplasty. Using a 126 subject dataset, machine learning models differentiated six tissue types with AUC from 0.79 to 0.63. The demonstrated radiomic approach shows promise for precise THA diagnostics and motivates further research using larger diverse data more sophisticated modeling techniques.

Conclusion

Conventional MRI radiomics can potentially reproduce expert classification of adverse soft tissue reactions following THA. In this work, a high-throughput quantitative radiomics-based approach demonstrated feasibility to diagnose subjective MRI interpretation for precision diagnostics. If rigorously validated, radiomics could become a powerful tool for guided management of THA complications, thus improving outcomes and reducing costs of THA management.