1256: A Machine Learning Pipeline for Prediction of Bone Marrow Oedema Along the Sacroiliac Joints on Magnetic Resonance Imaging

Joris Roels¹, Ann-Sophie Kathleen De Craemer², Thomas Renson³, Manouk de Hooge⁴, Arne Gevaert¹, Lennart Jans³, Nele Herregods³, Thomas Van Den Berghe³, Philippe Carron³, Filip Van den bosch⁵, Yvan Saeys¹ and Dirk Elewaut⁶, ¹Center for Inflammation Research, UGent-VIB, Ghent, Belgium, ²Ghent University Hospital, Gent, Belgium, ³Ghent University Hospital, Ghent, Belgium, ⁴Ghent University Hospital, Luxembourg, Luxembourg, ⁵Department of Internal Medicine and Paediatrics, Ghent University and VIB Centre for Inflammation Research, Ghent, Belgium, ⁶Department of Rheumatology, Ghent University Hospital, Belgium, VIB-UGent Center for Inflammation Research, Ghent University, Heusden, Belgium

Background/Purpose: Detection and semi-quantitative assessment of bone marrow oedema (BMO) on magnetic resonance imaging (MRI) of the sacroiliac (SI) joints is essential for the diagnosis and follow-up of spondyloarthritis (SpA) patients. A machine learning (ML) algorithm could address intra- and inter-reader variability and the time-consuming nature of these assessments.

Methods: We developed a deep learning workflow that automatically locates the SI joints, segments regions of interest (ilium and sacrum), performs quadrant extraction and predicts presence of BMO, suggestive of inflammatory lesions, on a quadrant-level in semi-coronal slices of T1/T2 weighted MRI scans (Figure 1). Using a ResNet18 backbone, the inflammation classifier was trained and tested (5-fold cross-validation) on scans of SpA patients (n=279), postpartum women (n=71), and healthy individuals (n=114); while 243 SpA MRIs served as external validation. Patient-level predictions were derived from aggregating quadrant-level predictions, i.e. at least one positive quadrant.

Results: The proposed method automatically detects the SI joints and segments the ilium and sacrum. An area under the receiver operating characteristics curve (ROC AUC) score of 94.7%, balanced accuracy (B-ACC) of 82.2% and an F1 score of 62.1% is reported for the proposed inflammation prediction algorithm in cross validation. In the external validation dataset, ROC AUC was 88.2%, B-ACC 72.1% and F1 score 50.8%. On a patient-level, the model achieved an accuracy of 81.6% and 81.4% in the cross validation and external validation dataset, respectively.

Conclusion: We propose a fully automated ML pipeline for detection of BMO on MRI of the SI joints that has the potential to assist clinicians in early diagnosis and reliable quantification of inflammatory lesions in patients with (suspected) SpA.
Figure 1: Proposed pipeline for automated scoring of the MRI data. The T1-weightes sequence is used for locating the SI joints and segmentation of the ilium/sacrum. Based on this, the SI regions of interest (ROI) are extracted, and (optionally) masked with the ilium/sacrum segments. These ROIs are then fed to a convolutional neural network to predict presence of (deep) inflammatory lesions.
Disclosures: J. Roels, None; A. De Craemer, None; T. Renson, None; M. de Hooge, None; A. Gevaert, None; L. Jans, None; N. Herregods, None; T. Van Den Berghe, None; P. Carron, Biogen, AbbVie/Abbott, Eli Lilly, Fresenius Kabi, Galapagos, Merck/MSD, Pfizer, Sanofi, UCB; F. Van den bosch, AbbVie, Lilly, Galapagos, Janssen, Merck, Novartis, Pfizer, UCB, Amgen, Bristol-Myers Squibb(BMS), Celgene; Y. Saeys, None; D. Elewaut, AbbVie, Eli Lilly, Galapagos, Novartis, UCB Pharma.