Skip to main content
ACR Convergence 2023 Main banner
Final Program


Schedule at a Glance
Legend
Session Format
In Person
In Person
Livestreamed
Livestreamed
On Demand
On Demand

Poster Session B

Spondyloarthritis (SpA) including psoriatic arthritis (PsA)

Session: (1383–1411) Spondyloarthritis Including Psoriatic Arthritis – Diagnosis, Manifestations, & Outcomes Poster II: Imaging & AS

1398: A Deep Learning Algorithm for MRI Spinal Inflammation in Axial Spondyloarthritis

Monday, November 13, 2023
9:00 AM - 11:00 AM PT
Location: Poster Hall
On Demand In Person

    Abstract Poster Presenter(s)

  • CC

    Chiu Wai Shirley Chan, MBBS

    Division of Rheumatology and Clinial Immunology, Department of Medicine, the University of Hong Kong
    Hong Kong, Hong Kong, China

    Disclosure information not submitted.

Shirley Chiu Wai Chan1, Ying Ying Lin2, Ho Yin Chung1, Kam Ho Lee3 and Peng Cao2, 1Divison of Rheumatology and Clinical Immunology, Department of Medicine, Queen Mary Hospital, the University of Hong Kong, Hong Kong, Hong Kong, 2Department of Diagnostic Radiology, the University of Hong Kong, Hong Kong, Hong Kong, 3Department of Radiology, Queen Mary Hospital, Hong Kong, Hong Kong, Hong Kong

Background/Purpose: Deep learning has achieved wide applications in different areas of medical imaging analysis. Limited studies have explored the use of deep learning for imaging interpretation in axial spondyloarthritis (SpA). This study aims to develop a deep learning algorithm for the detection of active spinal inflammation in short tau inversion recovery (STIR) sequence magnetic resonance imaging (MRI) in patients with axial SpA.

Methods: A total 330 patients with axial spondyloarthritis (SpA) were recruited. STIR MRI of the whole spine and clinical data were obtained. Spinal inflammation was defined by the presence of an active inflammatory lesion on the STIR sequence. Regions of interests (ROIs) were drawn outlining the active inflammatory lesion. 'Fake-color' images were constructed to mimic adjacent MRI slices (figure 1). Images from 270 patients and 60 patients were separated into the training/validation and testing sets, respectively (table 1). Deep learning algorithms were developed using attention-Unet. The performance of the neural network was compared to the image interpretation by a general radiologist blinded to the ground truth.

Results: Active inflammatory lesions were identified in 2891 MR images and were absent in 14590 MR images. The sensitivity and specificity of the derived deep neural network were 0.80±0.03 and 0.88±0.02, respectively. The Dice coefficient of the truth-positive lesions was 0.55±0.02. The area under the curve of the receiver operating characteristic (AUC-ROC) curve of the algorithm was 0.87±0.02. The performance of the developed deep neural network was comparable to the interpretation of a general radiologist with similar sensitivity and specificity (table 2).

Conclusion: A deep neural network was developed to detect spinal inflammation in axial SpA with performance comparable to a general radiologist.

Figure 1: The process of the generation of a 'fake-color' input image and the paired input label. The Blue arrow points to the 'fake-color' image. The orange arrow points to the original image. The blue box shows the zoom-in view.


Table 1: Baseline characteristics of the training and validation cohort.


Table 2: Sensitivity and specificity of deep neural network and radiologist in image level and individual spinal vertebral part.


S. Chan: None; Y. Lin: None; H. Chung: Eli Lilly, 6, GlaxoSmithKlein(GSK), 6, Pfizer, 6; K. Lee: None; P. Cao: None.