Unsupervised clustering of PET/CT features in fever of unknown origin (FUO) and inflammation of unknown origin (IUO).
Authors
Affiliations (4)
Affiliations (4)
- Department of Nuclear Medicine, University of Health Science Gaziosmanpasa Training and Research Hospital, Istanbul, Türkiye.
- Department of Infectious Diseases and Clinical Microbiology, Medical Faculty, Istanbul Medipol University, Istanbul, Türkiye.
- Department of Internal Medicine, Medical Faculty, Istanbul Medipol University, Istanbul, Türkiye.
- Department of Nuclear Medicine, Medical Faculty, Istanbul Medipol University, Istanbul, Türkiye.
Abstract
The aim of this study was to classify fever of unknown origin (FUO) patients based on PET/CT imaging features using unsupervised clustering, and to explore whether these patterns could provide insights into underlying etiologies. We conducted a retrospective cohort study of adult FUO patients who underwent 18F-FDG PET/CT. For each patient, data were extracted on involvement of reticuloendothelial organs, parenchymal organs, and large vessels, including the maximum SUVmax within each involved system and the uptake pattern (focal, diffuse, or mixed). Patients were clustered based on these PET/CT features using four unsupervised algorithms (hierarchical clustering, hierarchical density-based spatial clustering of applications with noise, spectral clustering, K-prototypes clustering), and a consensus clustering approach was applied to derive robust final clusters. Cluster assignments were compared with final clinical diagnoses using the Adjusted Rand Index (ARI) and Mutual Information (MI) to explore clinical relevance. The study included 289 FUO patients with a mean age of 59 ± 17 years, comprising 128 (44.3%) women and 161 (55.7%) men. Consensus clustering of PET/CT features in FUO patients identified four distinct clusters: Cluster 0 with combined reticuloendothelial, parenchymal, and large artery involvement, Cluster 1 with prominent reticuloendothelial and parenchymal involvement, Cluster 2 with mild reticuloendothelial involvement, and Cluster 3 with selective parenchymal involvement. Infections were the predominant etiologic category across all clusters, while rheumatologic diseases also contributed substantially, particularly in Cluster 2. Comparison with final diagnoses showed low concordance (ARI 0.006, NMI 0.047; Chi-square 32.46, <i>p</i> = 0.0012, Cramér's V = 0.19), indicating that consensus clusters capture PET/CT phenotypes rather than specific etiologies. Consensus clustering of PET/CT features in FUO patients revealed distinct imaging phenotypes, but these clusters were not effective in directly identifying final clinical diagnoses, underscoring PET/CT's complementary role in evaluation.