Learning from multiple readings for axial spondyloarthritis classification of the sacroiliac joints

TL;DR

This paper introduces an automated machine learning system to classify MRI lesions in the sacroiliac joints of axial spondyloarthritis patients, aiming to reduce variability and improve consistency.

Contribution

A novel multi-reader machine learning framework is proposed to model variability and automate lesion classification in axial spondyloarthritis MRI evaluations.

Findings

The automated system achieved AUCs between 0.85 and 0.99 for classifying five lesion types in sacroiliac joints.

Contouring accuracy showed 95% of test cases had errors below 2.76mm, comparable to expert inter-reader agreement.

Performance was validated on multiple clinical datasets including MEASURE-1, PREVENT, and SURPASS.

Abstract

Magnetic resonance imaging (MRI) is a cornerstone in the evaluation and monitoring of axial spondyloarthritis (axSpA), a chronic inflammatory condition primarily affecting the sacroiliac joints (SIJs), spine, entheses, and peripheral joints. Accurate quantification of axSpA-related changes on MRI is critical for effective research and patient management; however, current lesion detection and grading approaches suffer from substantial intra- and inter-reader variability, limiting their consistency and reliability. To address these challenges, we propose a fully automated machine learning system for SIJ delineation and lesion classification on coronal MRI. The end-to-end pipeline automatically extracts SIJ contours using a vector-field—based open-contour model and classifies the presence or absence of five lesion types (bone marrow oedema, ankylosis, sclerosis, erosions, and fatty lesions) using both T1-weighted and STIR sequences. A multi-reader learning framework is employed to explicitly model inter- and intra-reader variability by leveraging multiple readings and consensus labels. Model performance was evaluated using patient-wise cross-validation on data from the MEASURE-1 clinical trial and further validated on other clinical datasets (PREVENT, SURPASS). Lesion classification performance was assessed using area under the receiver operating characteristic curve (AUC), balanced accuracy, sensitivity, and specificity, while contouring accuracy was quantified using root-mean-square error, where we found that 95% of the whole test set had errors below 2.76mm. The proposed approach achieved AUCs ranging from 0.85 to 0.99 across the five lesion types, with the highest performance observed when using consensus-based labels, and results were comparable to expert inter-reader agreement. These findings demonstrate that fully automated SIJ delineation and lesion scoring can achieve expert-level performance and have the potential to reduce reader burden and variability in large-scale axSpA MRI studies.

Figures