Retrospective Case-Control Study for Developing an Artificial Intelligence (AI) Tool for Lesion Detection Using Magnetic Resonance Imaging (MRI) and Clinical Variables for Early Diagnosis of Axial Spondyloarthritis (axSpA)

BACKGROUND

The spondyloarthritis (SpA) are a group of chronic inflammatory diseases of autoimmune nature that share common clinical and genetic features, including an association with HLA-B27 antigen. They are among the most common rheumatic diseases with a prevalence of 0.01-2,5%. All of these conditions make the patients to move on a chronic disabling disease.

Patients with SpA can be classified based on their clinical presentation into either predominantly axial SpA (axSpA) or predominantly peripheral SpA. Axial SpA is characterized by primary involvement of the sacroiliac joints (SIJs) and/or the spine, leaading to substantial pain and disability. Until recently, the diagnosis of axSpA relied on detecting of structural changes evocative of sacroiliitis in the SIJs using plain radiography.

The introduction of Magnetic resonance imaging (MRI) for evaluating the SIJs has significantly advanced the recognition of axSpA. MRI can detect early inflammatory processes even in patients who do not yet have structural lesions. Besides, MRI has shown superiority over radiography in detecting structural changes in the SIJs. However, the definition of a "positive MRI" in SpA remains controversial, as both sensitivity and specificity have their limitations. Early diagnosis of SpA has become increasingly important, as treatments are now available, and MRI is emerging as the preferred choice for early diagnosis. A number of randomized controlled trials of anti-tumour necrosis factor agents in ankylosing spondylitis have demonstrated regression of inflammatory lesions in the spine by MRI. Moreover, the role of MRI in the early diagnosis of SpA has become better established, and imaging features of active sacroiliitis by MRI have been defined for axSpA diagnosis.

RATIONALE OF THE STUDY

Despite the current advances in medical imaging and ongoing efforts to improve the classification criteria for axSpA, a high proportion of axSpA patients remain under-diagnosed, leading to delays in diagnosis that can result in a poor prognosis. The volume of unstructured data coming from medical imaging contributes to diagnostic delays. The integration of AI and machine learning technologies in medicine for processing large datasets has led to faster and more accurate analysis, identification of real-world evidence gaps, and the agile generation of evidence to address healthcare providers' and healthcare systems' needs.

This study aims to develop an AI diagnostic tool that combines quantitative MRI data with clinical information to aid in the early diagnosis of axSpA.

OBJECTIVES

Primary objective: To create an AI tool that allows the early diagnosis of axSpA and lesion detection based on MRI.

Secondary objective: Clinical validation of the AI module.

Exploratory objective:

• Automated characterization of lesions (oedema, erosion, fat metaplasia and ankylosing) based on texture quantification and radiomics and deep features analysis.
• Determination of normative values for texture imaging biomarker on the SIJs.

SAMPLE DESCRIPTION

The dataset will consist of 900 MRIs, collected retrospectively. MRI exams will be sourced from patients with active axSpA and from those with inactive or no axSpA (control group). The control-to-case ratio will be set at 40/60, allowing the algorithm to learn from both subsets without favoring one group over the other. Since AI can more easily characterize normality than pathology, the proportion of non-axSpA and normal MRIs can be lower (approximately 40%) compared to the 60% allocated to axSpA MRIs. Among the active axSpA group, the distribution of MRIs across classification categories (oedema, ankylosing, erosion, and fat metaplasia) should be as balanced as possible, ideally with 25% assigned to each category. Each MRI does not necessarily come from a different patient, as they may represent different time points for the same individual.

ANALYSIS PLAN

1. Image Quality Control. All the images received from sites will be checked by imaging technicians to guarantee the homogeneity of the data
2. Centralized Image Interpretation. A centralized radiological review of the MRI images will be conducted by senior MSK expert radiologists. Each case will be evaluated by two radiologists. If there is a disagreement between the two, a third radiologist will review the case. The radiologists will classify the MRIs into the study's various classes and cohorts based on the ASAS criteria for defining active sacroiliitis on MRI for the classification of axial spondyloarthritis. All radiologists involved in the project will receive training to detect lesions according to the ASAS criteria, and this training will be documented and stored in the study's repository.
3. Annotation process. The imaging technicians will delineate the lesions detected by the MSK expert radiologists to generate a 3D volume. This will be then reviewed by the MSK expert radiologist.
4. Imaging Biomarkers Extraction. To obtain further information of the lesions labeled, a texture analysis will be performed to quantify several features related to the heterogeneity of the tissue that can be considered as an indicator of the pathological process. The radiomic panel will be based on the following features:

• Structural or shape features: Descriptive of the geometric properties of the image. Examples of these features are volume, maximum orthogonal diameter, maximum surface area, compactness, fractal dimension or sphericity of a lesion.
• Statistical characteristics are those that are inferred by statistical relationships. They can be in turn:
• First-order or distributional: They provide information on the frequency of individual voxel values without taking into consideration their spatial relationships. This distribution is presented in the form of histograms, which report the mean, median, maximum and minimum in the intensities of the voxels, but also on the asymmetry, kurtosis, uniformity or entropy of the distribution.
• Second order or texture: They reflect the relationships between neighboring voxels, allowing to obtain a spatial arrangement of their intensities, thus giving an idea about the architecture and heterogeneity of the studied tissue. These relationships are obtained by means of statistical analyses, such as cooccurrence matrices, which measure the probability that two neighboring voxels have the same signal intensity.
• Higher order: These are combinations of features obtained by complex statistical analysis, such as fractal analysis, on images to which filters or mathematical transformations have been applied to maximize or minimize patterns, remove noise, or highlight certain details.
• Deep features: These are properties obtained by analyzing images with convolutional neural networks (CNN) or other deep learning algorithms. These algorithms are trained to be able, in an image, to automatically determine and select those features or sets of classifying features, without the need of human intervention.

1. AI Module Development.

Using the MRIs collected together with the imaging biomarkers and other clinical information available, the data scientists will create an AI-based model that will provide a probability score of axSpA for each subject.

To create the AI module, the database will be divided in three non-balanced sub datasets (training, validation and test). The bigger dataset of images will be used for training the module. After the training phase, the validation database will be used to check if the classification is well done or the model should be trained again.

Training: Set of cases used to fit the model during the training process. These will be the cases the model will use to tune its weights, i.e. the cases that the model will use to "learn".

Validation: Set of cases used to evaluate the model during the training process, therefore, are not used to tune the model weights. This dataset is used for three main purposes:

• Hyperparameters tuning.
• Overfitting detection: During the training process, after each training iteration (epoch), the model is evaluated over both the training and validation datasets.
• Selection of the "best" model during the training process, this means that the weights from the training iteration in which the best validation metrics are obtained are stored.

Test: A separate set, not used for training or validation, will be used for the final model evaluation.

In this study, MRI exams will be obtained from various scanners and institutions. Therefore, the acquisition protocols and reconstruction techniques may vary between scanners. To address this, preprocessing techniques will be applied to standardise the images across different scanners.