AUTOMATIC DETECTION AND LOCALIZATION OF SPINAL LESIONS ON [18F]FLUORIDE PET/CT IN SPONDYLOARTHRITIS PATIENTS

Background: Spondyloarthritis (SpA) belongs to chronic inflammatory rheumatic diseases and primarily affects the axial skeleton. [ F]Fluoride PET/CT is an imaging approach that detects osteoblastic activity, and shows promise for early diagnosis and treatment assessment of SpA [1,2]. Visual assessment of [ F]Fluoride PET is time-consuming and subjective, and is ideally replaced by automatic methods. Objectives: This study aims to develop a fully automated lesion detection and localization algorithm combining detailed CT based segmentation of the axial skeleton with PET based lesion detection. Methods: 18 SpA patients who underwent [ F]Fluoride PET/CT at baseline, and 12 weeks and 52 weeks after the start of biological anti-rheumatic treatment, were included in this study. A nuclear medicine specialist and a musculoskeletal radiologist performed visual assessment of each anatomical structure, including the discovertebral units (DVUs), posterior spinal joints (costovertebral joints, costotransverse joints and facet joints) and sacroiliac joints (SIJs), and indicated whether there was positive [ F]Fluoride uptake. A multi-atlas based method, using initial deep learning based segmentation of the vertebrae on low-dose CT, was developed for the segmentation of the DVUs, posterior spinal joints and SIJs as volumes of interest (VOIs) [3,4] (Figure 1). For each VOI, the maximum, peak and mean standardized uptake values (SUVmax, SUVpeak and SUVmean) were extracted from the PET images. Next, for a range of SUV thresholds the PET uptake was classified as positive or negative and compared with visual assessment. Performance was evaluated by the area under the curve (AUC). To account for different background uptake between scans and patients, two methods of background corrections were applied, using the median SUV (SUVmedian) of the entire spine, and the SUVmedian for the cervical, thoracic and lumbar spine. The SUVmedian was obtained by using the initial segmentation of the vertebrae as VOIs. For the latter method, the VOIs were corrected by the corresponding SUVmedian dependent on their anatomical location, to account for differences in background uptake in the spinal regions. The differences between the AUC curves were assessed using the DeLong test. Results: 53 [ F]Fluoride PET/CT scans were analysed, which led to a total number of 6,254 VOIs, of which 567 (9.1%) were classified as [ F]Fluoride PET positive by the visual readers. The classification method that used SUVmax performed significantly better than the other metrics SUVpeak and SUVmean (AUC = 0.85 vs. 0.83; p <0.001, and 0.85 vs. 0.80; p <0.001, respectively) (Figure 2a). The maximum Kolmogorov-Smirnov (K-S) metric was achieved for SUVmax when a cut-off value of 10.4 was used. When the SUVmax in each scan was corrected for background uptake in the entire spine and the spinal regions, the classification performance improved significantly (AUC = 0.90 and AUC = 0.88 vs AUC = 0.85, respectively; both p <0.001). Correction by the SUVmedian spine was also significantly better than correction by the spinal regions (p <0.001) (Figure 2b). The optimal K-S metric in this corrected method was found at a cut-off value of 1.8 times the SUVmedian of the entire spine. Conclusion: This study presents an automatic algorithm for the detection and localization of [ F]Fluoride positive lesions in SpA patients, reaching high classification performance, with visual assessment as gold standard. SUVmax corrected by the SUVmedian of the entire spine is the best performing feature. Due to the relatively small prevalence of positive lesions, the cut-off values that give the highest K-S metric will result in a low positive predictive value of the classification method. The optimal cut-off value needs to be further investigated in future research projects, in which a lesion segmentation method will be developed. REFERENCES: [1] https://doi.org/10.1093/rheumatology/kex448 [2] https://doi.org/10.1007/s00259-022-06080-5 [3] https://doi.org/10.1148/ryai.230024 [4] https://doi.org/10.1038/s41592-020-01008-z Figure 1. Segmentation of DVUs (B), SIJs (C) and facet joints (D) Figure 2. Classification performance Acknowledgements: NIL. Disclosure of Interests: Wouter R.P. van der Heijden: None declared, Gerben J.C. Zwezerijnen: None declared, Floris H.P. van Velden: None declared, Ronald Boellaard received a research grant from Siemens Healthineers., Conny J. van der Laken worked as a consultant for GlaxoSmithKline GSK, UCB Pharma, Galapagos, Lilly, received research support from GlaxoSmithKline GSK, UCB Pharma, Novartis, Abbvie, Pfizer. DOI: 10.1136/annrheumdis-2024-eular.1230 Keywords: Artificial Intelligence, Imaging Citation: , volume 83, supplement 1, year 2024, page 887Session: Spondyloarthritis (Poster View)

Artificial Intelligence, Imaging