Akademik Veri Yönetim Sistemi
29th Annual Conference on Medical Image Understanding and Analysis, MIUA 2025, Leeds, İngiltere, 15 - 17 Temmuz 2025, cilt.15917 LNCS, ss.147-162, (Tam Metin Bildiri)
The diagnosis of radiographic axial spondyloarthritis (r-axSpA), including ankylosing spondylitis (AS), relies heavily on radiographic evaluation of the sacroiliac joints (SIJs) based on the modified New York (mNY) criteria. However, this process is hindered by low image quality, inter-reader variability, and difficulty in detecting early-stage bony changes. To address these limitations, this study proposes a novel image-processing-based diagnostic assistance tool designed to improve the classification and grading of SIJs in patients with axial spondyloarthritis (axSpA). The study involved consecutive axSpA patients under follow-up at Marmara University’s rheumatology outpatient clinics. Conventional SIJ radiographs were graded in a blinded manner by an experienced reader, with adjudication by a rheumatologist when necessary. The reference standard was definite sacroiliitis as defined by the mNY criteria. The proposed model utilizes a two-phase pipeline: the first phase employs the VGG-16 deep convolutional neural network for binary classification of r-axSpA and non-radiographic axSpA (nr-axSpA), while the second phase performs multiclass grading of sacroiliitis severity. The binary classification achieved an accuracy of 91.98%, and the multiclass grading phase reached 93.01% accuracy. Model performance was evaluated using the Confusion Matrix and F1 Score, highlighting its strong alignment with expert readings and robustness against false positives and negatives. This image-based approach demonstrates the potential to assist clinicians in the diagnosis of axSpA, offering a more objective and reproducible alternative to traditional visual grading. It supports clinical decision-making by enhancing diagnostic precision and facilitating the early identification of r-axSpA.