¹Trakia University, Department of Health Care, Faculty of Medicine, Stara Zagora
²University Prof Dr Assen Zlatarov, Faculty of Global Health and Health Care, Burgas
³Heidelberg University Hospital, Institute of Medical Informatics, Heidelberg
⁴Ruhr University Bochum, and Rheumazentrum Ruhrgebiet, Herne

Text

Introduction: Axial spondyloarthritis (axSpA) is a chronic inflammatory disease that can lead to significant spinal deformities, particularly kyphosis [1]. Accurate measurement of the kyphosis angle is crucial for monitoring disease progression and treatment response [2]. Traditionally, the Cobb angle, which measures the degree of spinal curvature, is manually calculated by identifying the uppermost and lowermost vertebrae within the thoracic spine [3]. However, this method can be time-consuming, subjective and prone to human errors. There is a need for automated and reliable methods to assess spinal curvature, particularly in clinical environments where high-throughput image analysis is required [4]. This study aims to leverage the You Only Look Once (YOLO) object detection model for the automated detection of vertebrae in morphometric analyses of Dual-Energy X-ray Absorptiometry (DEXA) spine images and to calculate the kyphosis angle. Furthermore, our goal was to compare the kyphosis angle estimated by the YOLO model and the Cobb angle measured by physicians.

Methods: In this study, we trained a YOLO model to detect vertebrae in morphometric analyses of DEXA images. The dataset comprised 512 annotated DEXA images, including 182 images of axSpA patients and 330 images from other patients who underwent DEXA scans for various reasons. These images were manually labeled to identify vertebral bodies and their corresponding bounding boxes. After training, the YOLO model predicted the centers of the vertebrae, which were subsequently used to fit a circle, approximating the curvature of the spine. To calculate the kyphosis angle, the positions of the uppermost (Th4) and lowermost (Th12) vertebrae were extracted. The kyphosis angle was then determined by measuring the angle formed between these two vertebrae and the radius of the fitted circle. Finally, the computed kyphosis angles from the YOLO model were compared to Cobb angles manually measured by physicians. Statistical analysis was performed to assess the agreement between the model’s estimated kyphosis angles and the Cobb angle measurements.

Results: The mean kyphosis angle of axSpA patients estimated by the YOLO model was 45.3° ± 7.8° (SD, range 28°–75°), and by physicians it was 46.1° ± 8.2° (range 30°–78°). The mean kyphosis angle of other patients estimated by the YOLO model was 35.2° ± 6.5° (SD, range 20°–60°), and by physicians it was 36.5° ± 7.0° (range 21°–62°).The model showed a strong correlation with the measurements of the physicians, with a Pearson correlation coefficient of 0.92 (p < 0.001). The mean squared error (MSE) between the model’s computed kyphosis angles and the Cobb angle was 4.2°, indicating good agreement between the automated method and the manual physician measurements. The sensitivity and specificity of the model in detecting clinically significant kyphosis (defined as a Cobb angle ≥ 40°) were 0.89 and 0.91, respectively. These results demonstrate that the YOLO-based model is both accurate and reliable in estimating kyphosis angles.

Conclusion: This study demonstrates the feasibility of using a YOLO-based deep learning approach for detecting vertebrae in morphometric analyses of DEXA images and calculating the kyphosis angle in axSpA patients. The model showed a high degree of accuracy in estimating kyphosis angles with results closely aligning with the Cobb angle measured by physicians. This automated approach offers a promising alternative to traditional manual methods, providing a faster, more reproducible and potentially more accurate tool for spinal deformity assessment in clinical settings. Future work will focus on refining the model with larger, more diverse datasets and integrating it into clinical decision support systems for improved patient care and monitoring.

Literatur

[1] Walsh JA, Magrey M. Clinical Manifestations and Diagnosis of Axial Spondyloarthritis. J Clin Rheumatol. 2021 Dec 1;27(8):e547-e560. DOI: 10.1097/RHU.0000000000001575
[2] Garrido-Castro JL, Ladehesa Pineda ML, López-Medina C, et al. POS0948 Hyperkyphosis in axial spondyloarthritis: Factors that produce its appearance and effect on mobility and function. Annals of the Rheumatic Diseases. 2021;80(Suppl 1):737-8. DOI: 10.1136/annrheumdis-2021-eular.346
[3] Kado DM, Prenovost K, Crandall C. Narrative review: hyperkyphosis in older persons. Ann Intern Med. 2007 Sep 4;147(5):330-8. DOI: 10.7326/0003-4819-147-5-200709040-00008
[4] Joo YB, Baek IW, Park KS, Tagkopoulos I, Kim KJ. Novel classification of axial spondyloarthritis to predict radiographic progression using machine learning. Clin Exp Rheumatol. 2021 May-Jun;39(3):508-18. DOI: 10.55563/clinexprheumatol/217pmi