https://orcid.org/0000-0001-8310-7820
Abstract
To evaluate whether neural networks can distinguish between seropositive RA, seronegative RA, and PsA based on inflammatory patterns from hand MRIs and to test how psoriasis patients with subclinical inflammation fit into such patterns.
ResNet neural networks were utilized to compare seropositive RA vs PsA, seronegative RA vs PsA, and seropositive vs seronegative RA with respect to hand MRI data. Results from T1 coronal, T2 coronal, T1 coronal and axial fat-suppressed contrast-enhanced (CE), and T2 fat-suppressed axial sequences were used. The performance of such trained networks was analysed by the area under the receiver operating characteristics curve (AUROC) with and without presentation of demographic and clinical parameters. Additionally, the trained networks were applied to psoriasis patients without clinical arthritis.
MRI scans from 649 patients (135 seronegative RA, 190 seropositive RA, 177 PsA, 147 psoriasis) were fed into ResNet neural networks. The AUROC was 75% for seropositive RA vs PsA, 74% for seronegative RA vs PsA, and 67% for seropositive vs seronegative RA. All MRI sequences were relevant for classification, however, when deleting contrast agent–based sequences the loss of performance was only marginal. The addition of demographic and clinical data to the networks did not provide significant improvements for classification. Psoriasis patients were mostly assigned to PsA by the neural networks, suggesting that a PsA-like MRI pattern may be present early in the course of psoriatic disease.
Neural networks can be successfully trained to distinguish MRI inflammation related to seropositive RA, seronegative RA, and PsA.
Distinguishing different forms of arthritis using MRI and neural networks is possible.
Established MRI sequences are relevant for classification, but removing contrast-enhanced sequences did not decrease performance.
Most psoriasis cases were assigned to PsA, suggesting PsA-like MRI patterns in early psoriasis.
Introduction
The hand is one of the most prominent sites for inflammatory arthritis. RA and PsA both affect the hand and lead to inflammatory and structural lesions in the hand joints [1, 2]. MRI is widely used to detect and monitor inflammation in the hand joints and specific instruments have been developed and validated that allow the quantification of inflammatory lesions such as synovitis, tenosynovitis, and osteitis in the hand joints of RA and PsA patients [3, 4].
While MRI evaluation of joints has been primarily used to quantify inflammation, less is known about the potential of MRI in distinguishing different patterns of inflammation in the various forms of arthritis. It is known that the imaging findings between these different disease entities are overlapping and are commonly non-specific [5]. Therefore it is challenging to define a specific MRI pattern that allows differentiation between different forms of arthritides. In order to extract such MRI patterns differentiating seropositive RA, seronegative RA and PsA, a suitable neural network is required. A popular choice for the classification of medical images are residual networks (ResNets) [6, 7].
In this study, we assessed hand MRI data by a ResNet with the aim of testing whether such a neural network is able to differentiate seropositive RA, seronegative RA and PsA based on hand MRI data, assessing which MR sequences are relevant for classification and testing whether MRI data from psoriasis patients with subclinical inflammation fall into an RA-like or PsA-like pattern.
Methods
Patients
Patients were part of the Erlangen Imaging Cohort [8]. MRI examination of the hand acquired between 2018 and 2020 were evaluated. RA patients fulfilled 2010 EULAR/ACR criteria, PsA patients fulfilled the Classification Criteria for Psoriatic Arthritis. Psoriasis patients were recruited at the Dermatology Department and referred to the Department of Rheumatology for clinical examination for signs of musculoskeletal involvement (pain, swelling, enthesitis, dactylitis, inflammatory back pain) and for MRI evaluation. Psoriasis patients had no signs of PsA. Additional clinical follow-up for psoriasis patients was done to assess progression to PsA, as described previously [9]. In all patients, demographic parameters, disease activity parameters and functional parameters were recorded. Ethical approval (334_18B; 152_20B) was obtained from the local ethics committee of the University Clinic Erlangen and all patients gave written informed consent.
MRI
MRIs of the dominant hand were acquired on 1.5 T Magnetom Avanto and Aera systems (Siemens, Erlangen, Germany). The full MRI protocol consisted of five sequences: coronal T1, coronal T1 fat-suppressed contrast enhanced (CE), coronal T2 fat-suppressed, axial T2 fat-suppressed and axial T1 fat-suppressed CE (see also the supplementary material, available at Rheumatology online).
Deep learning
In the experiments, ResNets pretrained on video understanding were used [10]. Each MRI sequence was processed by a separate network before the information was fused using an ensemble of all individual predictions. In a second step, demographic and clinical features (patient’s weight, height, sex, smoking status, CRP, swollen joint count, tender joint count, HAQ, 28-joint DAS) was utilized using a separate network, after which the prediction is formed using an ensemble (Fig. 1). MRI scans were individually normalized to have zero mean and unit variance and were reshaped to the same size (512 × 512 × 16 voxels for coronal scans, 320 × 320 × 64 for axial scans). Further, RandAugment [16] was applied as an augmentation strategy in the training phase. We considered only the binary classification of the three classes rather than a three-class classification, resulting in three experiments (seropositive RA vs PsA, seronegative RA vs PsA, seronegative vs seropositive RA). This approach was based on the assumption that seropositive and seronegative RA are difficult to distinguish on MRI and also present similarly from a clinical perspective [11] and an early deep-learning experiment (Supplementary Table S1, available at Rheumatology online) demonstrated that the three-class classification did not allow us to distinguish well between different forms of arthritis due to poor discrimination between seronegative and seropositive RA cases.
Overview of MRI and clinical data integration for deep learning analysis
Each MRI sequence (T-weighted coronal and T1-weighted fat-suppressed contrast-enhanced axial sequences are depicted as examples) is processed by an individual neural network (ResNet 3D). Clinical data are optionally added to each case and processed similarly by a further neural network. The final prediction for a single case is calculated by average over all individual network predictions. Images adapted from https://www.philips.co.uk/healthcare/education-resources/publications/fieldstrength/expanding-the-imaging-center#hand_and_wrist_mri
To explore which sequences and combination of sequences are most relevant for the network, we compared networks with different combinations of sequences as input. We used 5-fold cross-validation, which resulted in a split of training and test data of 80% and 20%, respectively. Finally, the trained networks were applied to psoriasis patients to gain a deeper understanding of their similarities to PsA and RA.
All experiments were executed on a V100 GPU (Nvidia, Santa Clara, CA, USA) using PyTorch 1.7 (https://pytorch.org/) on Python 3.8 (https://www.python.org/) and SciPy 1.7 (https://scipy.org/).
Impact of anatomical regions
In order to increase the understanding for the prediction of a neural network, we utilized recent visualization techniques [12]. By repeatedly occluding patches of the input image with empty masks, the perturbation to the classification network prediction was measured indicating regions relevant to the task. In this exercise, repeated predictions are made using the trained network on MRIs after a different region in the image is coded as missing at each iteration. The difference of predicted probabilities between those from full images and such partially missing images is used to quantify the importance of the missing region for classification.
Impact of the size of training data sets
The amount of training data necessary for a neural network with good generalization and performance highly depends on the task investigated [13]. Thus we analysed the performance of varying amounts of training data sets in terms of area under the receiver operating characteristic curve (AUROC) while keeping the test set fixed.
Reader studies
Two reader studies were performed on a randomly selected subset of the data set (process described in statistics section). The aim of the first reader study was to determine whether an experienced MRI reader could achieve comparable classification performance of seropositive RA and PsA compared with the neural network, based on systematic imaging evaluation of the same MRI sequences and with additional use of the same demographic and clinical parameters provided to the neural network. For this purpose, the reader first assessed MRIs using the Psoriatic Arthritis MRI Score (PSAMRIS)/Rheumatoid Arthritis MRI Score (RAMRIS) systems and then utilized the scoring results and the demographic/clinical data for classifying each case into seropositive RA, PsA or ‘diagnosis not possible’.
The aim of the second reader study was to determine whether the anatomical regions considered most important by the neural network matched those considered essential by an experienced rheumatologist. For this study, the PSAMRIS/RAMRIS results of the hand MRI from the first reader were used. Then a second reader evaluated heat maps generated by the neural network using Grad-CAM images (http://gradcam.cloudcv.org/). The same anatomical regions as assessed by the first reader were used to classify whether the respective anatomical region received high attention from the network-generated heat map. The PSAMRIS/RAMRIS scores obtained by the first reader were then compared with the scores obtained by the second reader using Spearman rank-order correlation. Data from a subset of 22 patients from the seropositive RA and PsA groups with all MRI sequences were used. The expert readers were blinded to each other and to the patient data.
Statistics
We described participant characteristics using summary statistics for continuous and categorical data. Statistical significance was determined using a one-sided Wilcoxon, one-sample rank sum test on the AUROC using the test sets in the cross-validation. The s.d. of the AUROCs was calculated during a 5-fold stratified cross-validation. The underlying null hypothesis was that a given AUROC is equal to or less than the reference AUROC (i.e. 0.5). Statistical significance was determined at α = 0.05. When computing the metrics, PsA was considered the positive class while RA was considered the negative class. For the comparison of seronegative and seropositive RA, the latter was considered the negative class and the former as the positive class. Binary metrics were calculated using a threshold of 0.5. For the correlation analysis between the rheumatologist and the regions identified by the networks described in the reader study, Spearman rank-order correlation was performed. The samples for the reader study were drawn in the following way: In a 5-fold cross-validation manner each test case was embedded using all but the last layer of the classification network. Then, a t-distributed stochastic neighbour embedding of the feature vectors to a two-dimensional space was performed (Supplementary Figure S1, available at Rheumatology online) to yield cases evenly distributed over the feature space. Finally, the cases were sampled in a uniform random manner from that low-dimensional space.
Results
Patients’ characteristics
In total, MRI scans of 649 patients were analysed (382 women/267 men). The patients’ characteristics are depicted in Table 1. The mean age of the patients was 55.4 years (s.d. 13.0). A total of 135 patients had seronegative RA, 190 had seropositive RA, 177 had PsA and 147 had psoriasis without clinical joint disease. The mean disease activity was moderate and comparable among the three forms of arthritis. Of the psoriasis patients, 14 developed PsA over time.
Overview of demographic and clinical information of the patient cohort (N = 649)
| Characteristics | Seropositive RA | Seronegative RA | PsA | Psoriasis |
|---|---|---|---|---|
| Patients, n | 190 | 135 | 177 | 147 |
| Age, years, mean (s.d.) | 56.9 (12.6) | 60.5 (10.3) | 56.3 (12.0) | 49.6 (13.8) |
| Female/male, n | 126/64 | 93/42 | 92/85 | 71/76 |
| BMI, kg/m2, mean (s.d.) | 26.6 (10.5) | 27.6 (9.3) | 29.1 (11.3) | 26.7 (6.9) |
| Disease duration, years, mean (s.d.) | 2.6 (4.9) | 1.3 (2.3) | 0.8 (2.3) | 4.2 (5.1) |
| DAS28, mean (s.d.) | 3.3 (1.3) | 3.4 (1.2) | 3.2 (1.3) | – |
| CRP, mg/L, mean (s.d.) | 0.9 (2.5) | 0.7 (1.2) | 0.5 (0.8) | 0.5 (1.3) |
| HAQ, mean (s.d.) | 0.8 (0.6) | 0.9 (0.8) | 0.6 (0.6) | 0.3 (0.4) |
| Medication, n (%) | ||||
| bDMARD | 168 (88.4) | 113 (83.8) | 144 (81.3) | 51 (35.0) |
| csDMARD | 170 (89.5) | 120 (88.8) | 142 (80.5) | 18 (12.2) |
| Characteristics | Seropositive RA | Seronegative RA | PsA | Psoriasis |
|---|---|---|---|---|
| Patients, n | 190 | 135 | 177 | 147 |
| Age, years, mean (s.d.) | 56.9 (12.6) | 60.5 (10.3) | 56.3 (12.0) | 49.6 (13.8) |
| Female/male, n | 126/64 | 93/42 | 92/85 | 71/76 |
| BMI, kg/m2, mean (s.d.) | 26.6 (10.5) | 27.6 (9.3) | 29.1 (11.3) | 26.7 (6.9) |
| Disease duration, years, mean (s.d.) | 2.6 (4.9) | 1.3 (2.3) | 0.8 (2.3) | 4.2 (5.1) |
| DAS28, mean (s.d.) | 3.3 (1.3) | 3.4 (1.2) | 3.2 (1.3) | – |
| CRP, mg/L, mean (s.d.) | 0.9 (2.5) | 0.7 (1.2) | 0.5 (0.8) | 0.5 (1.3) |
| HAQ, mean (s.d.) | 0.8 (0.6) | 0.9 (0.8) | 0.6 (0.6) | 0.3 (0.4) |
| Medication, n (%) | ||||
| bDMARD | 168 (88.4) | 113 (83.8) | 144 (81.3) | 51 (35.0) |
| csDMARD | 170 (89.5) | 120 (88.8) | 142 (80.5) | 18 (12.2) |
bDMARD: biologic DMARD; csDMARD: conventional synthetic DMARD.
Overview of demographic and clinical information of the patient cohort (N = 649)
| Characteristics | Seropositive RA | Seronegative RA | PsA | Psoriasis |
|---|---|---|---|---|
| Patients, n | 190 | 135 | 177 | 147 |
| Age, years, mean (s.d.) | 56.9 (12.6) | 60.5 (10.3) | 56.3 (12.0) | 49.6 (13.8) |
| Female/male, n | 126/64 | 93/42 | 92/85 | 71/76 |
| BMI, kg/m2, mean (s.d.) | 26.6 (10.5) | 27.6 (9.3) | 29.1 (11.3) | 26.7 (6.9) |
| Disease duration, years, mean (s.d.) | 2.6 (4.9) | 1.3 (2.3) | 0.8 (2.3) | 4.2 (5.1) |
| DAS28, mean (s.d.) | 3.3 (1.3) | 3.4 (1.2) | 3.2 (1.3) | – |
| CRP, mg/L, mean (s.d.) | 0.9 (2.5) | 0.7 (1.2) | 0.5 (0.8) | 0.5 (1.3) |
| HAQ, mean (s.d.) | 0.8 (0.6) | 0.9 (0.8) | 0.6 (0.6) | 0.3 (0.4) |
| Medication, n (%) | ||||
| bDMARD | 168 (88.4) | 113 (83.8) | 144 (81.3) | 51 (35.0) |
| csDMARD | 170 (89.5) | 120 (88.8) | 142 (80.5) | 18 (12.2) |
| Characteristics | Seropositive RA | Seronegative RA | PsA | Psoriasis |
|---|---|---|---|---|
| Patients, n | 190 | 135 | 177 | 147 |
| Age, years, mean (s.d.) | 56.9 (12.6) | 60.5 (10.3) | 56.3 (12.0) | 49.6 (13.8) |
| Female/male, n | 126/64 | 93/42 | 92/85 | 71/76 |
| BMI, kg/m2, mean (s.d.) | 26.6 (10.5) | 27.6 (9.3) | 29.1 (11.3) | 26.7 (6.9) |
| Disease duration, years, mean (s.d.) | 2.6 (4.9) | 1.3 (2.3) | 0.8 (2.3) | 4.2 (5.1) |
| DAS28, mean (s.d.) | 3.3 (1.3) | 3.4 (1.2) | 3.2 (1.3) | – |
| CRP, mg/L, mean (s.d.) | 0.9 (2.5) | 0.7 (1.2) | 0.5 (0.8) | 0.5 (1.3) |
| HAQ, mean (s.d.) | 0.8 (0.6) | 0.9 (0.8) | 0.6 (0.6) | 0.3 (0.4) |
| Medication, n (%) | ||||
| bDMARD | 168 (88.4) | 113 (83.8) | 144 (81.3) | 51 (35.0) |
| csDMARD | 170 (89.5) | 120 (88.8) | 142 (80.5) | 18 (12.2) |
bDMARD: biologic DMARD; csDMARD: conventional synthetic DMARD.
The four groups were equally distributed between the two MRI systems [Aera scanner: seropositive RA, N = 147 (77.3%); seronegative RA, N = 107 (79.1%); PsA, N = 137 (77.6%); psoriasis, N = 123 (83.4%)].
Differentiation between the different forms of arthritis
We first examined the differentiation between seronegative RA and PsA based on the information extracted by the network using different combinations of the five MR sequences without further clinical information. The highest AUROC with 74% was achieved by the model combining T1 coronal CE, T2 coronal, T1 axial CE, and T2 axial sequences. The corresponding accuracy, sensitivity and specificity were 65%, 75% and 58%, respectively. The combination of only two sequences, T1 coronal CE and T2 axial, achieved an AUROC of 73% (overview of all combinations in Supplementary Table S2, available at Rheumatology online). The AUROC of all configurations in the performed experiments was significantly >0.5.
When testing the differentiation between seropositive RA and PsA, the combination of all but the T1 coronal CE sequence yielded an AUROC of 75%, with an accuracy of 68%, sensitivity of 80% and specificity of 54%. A protocol consisting of native T1 coronal, T2 coronal and T2 axial sequences, leaving out contrast-enhanced sequences, yielded an AUROC of 73% (overview of all combinations in Supplementary Table S3, available at Rheumatology online). When differentiating seropositive and seronegative RA by the neural network, the best results could be achieved by combining T1 coronal with and without contrast agent with an AUROC of 67%; accuracy was 62%, sensitivity 69% and specificity 53% (Supplementary Table S4, available at Rheumatology online).
Integration of clinical and serological parameters
In addition, additive integration of clinical data into the network was explored regarding their possibility to improve differentiation between PsA and seropositive RA. The best performance could be achieved by the combination of all sequences (except T1 axial CE), with an AUROC of 75%, an accuracy of 68%, a sensitivity of 80% and a specificity of 54% between PsA and seropositive RA (Supplementary Table S5, available at Rheumatology online).
The best performance for the differentiation of seronegative RA and PsA with additional clinical data was achieved using all but the T1 coronal sequence, with an AUROC of 74%, a sensitivity of 75% and a specificity of 58% (Supplementary Table S6, available at Rheumatology online). Finally, for the comparison of seronegative RA and seropositive RA, the best configuration was achieved by combining the T1 coronal CE and T2 coronal sequence, with an AUROC of 66%, a sensitivity of 69% and a specificity of 53% (Supplementary Table S7, available at Rheumatology online).
Clinical data did not add to better differentiation as compared with the network operating only on the MRI sequences.
Application of the trained neural networks in psoriasis patients
To further understand the mechanisms behind the network, we applied both trained networks to psoriasis patients and analysed to which form of inflammatory joint disease these patients cluster. In the first network, which differentiated seropositive RA from PsA, MRIs of 145 of the 147 psoriasis patients (98.6%) assigned a higher probability for PsA, while only 2 of the 147 psoriasis patients (1.4%) were assigned to seropositive RA. For all 14 psoriasis patients that later developed PsA, the network assigned the MRI to ‘PsA’. For the second network, which differentiated seronegative RA from PsA, 47 psoriasis cases (32.0%) were classified as seronegative RA and 100 cases (68.0%) as PsA. For 11 of the 14 patients developing PsA, the network predicted PsA.
Impact of training data size and anatomical structures to neural networks
Increasing the number of patients used to train the neural network had a substantial influence on the predictions for the test set (Fig. 2). With 20% of the available training data an AUROC of 65% (s.d. 6) was achieved, with 40% of the training data an AUROC of 70% (s.d. 5) was achieved and with the whole data set an AUROC of 74% (s.d. 8) was obtained. The results of the first reading study showed that in 50% of the 22 test cases, 10 of 11 patients were correctly classified (6 of 6 RA, 4 of 5 PsA) (91% accuracy, 90% sensitivity, 93% specificity and 92% F1), but in the remaining 50% of cases (9 RA, 2 PsA), no clear diagnosis was possible. In the second reader study it was found that the neural network considered some regions (especially MCP joints) that also had clinical relevance (Fig. 3, Supplementary Table S8, available at Rheumatology online).
Influence of training set sizes on the performance of the neuronal network
AUROC (y-axis) with different amounts of training data (x-axis; 0.2 = 20% of training data; 1.0 = 100% of training data). The performance was assessed five times with different samples showing mean performance (dark blue line) and variance (light blue area).
Visualization of heat maps generated by the neuronal network
(A and B) Neural network–generated heat maps overlapped with T2-weighted MRIs of the hand of patients with RA and PsA, respectively. Asterisks, arrows and arrowheads mark regions of pathological joint changes that correspond to artificial intelligence–detected hotspots. (A) Hand of a patient affected by RA with carpal and MCP2 synovitis (white arrows), MCP3 osteitis (asterisk) and an early erosion of the third metacarpal head (white arrowheads). (B) Hand of a patient affected by PsA with radiocarpal and ulnocarpal synovitis (white arrows), osteitis of the MCP3, PIP1 and PIP 2 (asterisks); extracapsular soft tissue inflammation and enthesitis of the joint capsule, extensor tendons and collateral ligaments (white arrowheads).
Discussion
These data show that neural network–based deep learning can identify patterns of inflammatory lesions in hand MRIs that may allow differentiation of seropositive RA, seronegative RA and PsA. The addition of clinical data had little impact on improving this performance, suggesting that the main forms of inflammatory arthritides differ in their distribution pattern of inflammation. Recently a small study of <50 patients aimed to use machine learning methods for the classification of inflammatory arthritis, demonstrating the potential of such computational approaches, however, this study still relied significantly on human experts to identify relevant structures [14]. So far there has been no study that has exclusively used MRI data and deep learning without requiring further expert input for the classification of arthritides.
Deep learning–based interpretation of MRI patterns may support classifying unclear cases of inflammatory arthritis, in particular distinguishing PsA from seronegative RA. The absence of psoriasis in PsA and the absence of diagnostic antibodies in RA can make such classification challenging and additional instruments are needed to improve this differentiation. The non-invasive character of MRI paired with the possibility of deep learning–based interpretation of the pattern of inflammatory lesions make this approach attractive with clinically more challenging cases. Of note, in the reader study we were able to demonstrate these challenges, as the visual evaluation and classification of MRI data failed to reach a clear decision in half of cases, likely due to the non-specificity of lesions. Hence MRI data could be used as a decision support tool for radiologists and rheumatologists in the future.
The data also showed that for distinction of RA and PsA, the applied algorithm did not necessarily require all MRI sequences. Thus less than half of the MRI sequences of the protocol was required to achieve almost comparable performance. Interestingly, axial T2-weighted fat-suppressed sequences alone also yielded considerably good results, suggesting that specific sequences may allow better pattern recognition by neural networks. One could speculate that in the future such an algorithmic approach could lead to an adaption of diagnostic MRI protocols with a smaller number of sequences and omission of contrast agent administration, which sometimes can induce toxicity [15]. However, further studies need to be conducted to confirm these results. From a clinical perspective, two additional findings are interesting. First, the overlap between the pattern of MRI changes between seropositive and seronegative RA is more substantial than between both forms of RA and PsA. Second, the classification of most psoriasis cases, including all psoriasis cases that developed PsA as PsA rather than RA, suggests that there might be early changes or other structural peculiarities in psoriasis patients recognized by the neural network leading to such classification, with patterns not fitting those of RA. We will need to refine our network application in larger cohorts of psoriasis patients as well as other subjects without hand arthritis to fully understand the characteristics of this classification behaviour.
Deep learning–based analysis of MRI patterns of arthritis still has considerable limitations. Classification accuracy was only moderate, which can be explained by the fact that the relatively large data set of hand MRIs from >600 patients was still not sufficiently large, as indicated by the unsaturated training curve. However, this also indicates that the method still has room to improve classification accuracy. A lack of controls for the psoriasis classification task is a limitation to explaining the overwhelming majority of psoriasis patients being classified into the PsA category. Our work can be considered preliminary and not fully applicable to practice at this stage. However, the primary focus of our analysis was assessing the capability of the deep learning approach for MRI patterns of arthritis. Thus the applicability for more complex classifications or for distinguishing between the presence/absence of inflammation or disease merits further investigation.
In summary, our results suggest the existence of MRI changes that the neural network identified as relevant for the classification of different forms of arthritis that have not been described before. These results indicate that the accuracy can be improved with further model training and can potentially achieve a level to allow clinical applicability. Future analysis of the regions highlighted by the network could lead to a better understanding of disease mechanisms.
Acknowledgements
Infrastructure and hardware support was provided by the d.hip Digital Health Innovation Platform.
Funding: The study was supported by the Deutsche Forschungsgemeinschaft (DFG-FOR2886 PANDORA and the CRC1181 Checkpoints for Resolution of Inflammation). Additional funding was received by the Bundesministerium für Bildung und Forschung (project MASCARA), the ERC Synergy grant 4D Nanoscope, the IMI funded projects HIPPOCRATES and RTCure, the Emerging Fields Initiative MIRACLE of the Friedrich-Alexander-Universität Erlangen-Nürnberg and the Else Kröner-Memorial Scholarship (to D.S.; no. 2019_EKMS.27).
Disclosure statement: The authors have declared no conflicts of interest.
Data availability statement
The software will be published on GitHub alongside this manuscript: https://github.com/lukasfolle/MRI-Classification-RA-PsA. We agree to make data supporting the results and analyses presented in this article available upon reasonable request.
Supplementary data
Supplementary data are available at Rheumatology online.
References
Author notes
Andreas Maier, Frank Roemer and David Simon contributed equally to this study.
Comments