1393: Fully Automated Detection of Active Sacroiliitis in Patients with Axial Spondyloarthritis: A Machine Learning-Based Analysis Magnetic Resonance Image

Go-Eun Lee¹, Sang-Il Choi¹, Jungchan Cho², Seon Ho Kim³, Geun Young Lee⁴ and Sang Tae Choi⁵, ¹Department of Computer Engineering, Dankook University, Seongnam, South Korea, ²School of Computing, Gachon University, Seongnam, South Korea, ³Integrated Media Systems Center, Viterbi School of Engineering, University of Southern California, Los Angeles, CA, ⁴Department of Radiology, Chung-Ang University College of Medicine, Seoul, South Korea, ⁵Division of Rheumatology, Department of Internal Medicine, Chung-Ang University College of Medicine, Seoul, South Korea

Background/Purpose: Magnetic Resonance Imaging (MRI) is a crucial modality for early diagnosis of active inflammation in the sacroiliac joint in patients with axial spondyloarthritis (axSpA). This study focused on developing a fully automated classification model that leverages machine learning to detect sacroiliac joints and determine the presence of bone marrow edema in MRI.

Methods: We collected 815 MRI slices of sacroiliac joints (SIJs) from 60 axSpA patients and 19 healthy subjects. Active sacroiliitis was identified by bone marrow edema observed in gadolinium-enhanced fat-suppressed T1-weighted oblique coronal images. First, a region of interest (ROI) was manually set, and the ResNet18 model was applied to detect bone marrow edema automatically. The prediction models were evaluated using 5-fold cross-validation sets. In the second phase, we introduced a text-guided cross-position attention module (CPAMTG) that integrates cross-attention into the position attention module (PAM) to localize the ROI automatically. The effectiveness of attention in extracting feature maps was assessed by comparison with backbone networks (U-Net) and PAM.

Results: The semi-automated model demonstrated commendable performance in detecting bone marrow edema, with 77.48% accuracy, 92.15% recall, 73.43% precision, 74.24% specificity, and an F1 score of 81.74% at the image level. At the patient level, active sacroiliitis was diagnosed with 96.06% accuracy, 100% recall, 94.84% precision, 86.43% specificity, and an F1 score of 97.32%. Remarkably, the fully automated ROI patch exhibited higher accuracy (84.73% vs. 77.48%, p < 0.001) and specificity (85.03% vs. 74.24%, p = 0.004) and maintained or improved performance in comparison to the semi-automated model, with 92.17% recall, 82.81% precision, and an F1 score of 87.24%.

Conclusion: We presented a fully automated classification model for detecting active sacroiliitis in MRI, which showed excellent performance. These findings suggest that MRI analysis with machine learning can offer valuable assistance to clinicians, enabling rapid and objective diagnosis of active inflammation in patients with axSpA.


G. Lee: None; S. Choi: None; J. Cho: None; S. Kim: None; G. Lee: None; S. Choi: None.