FUO-PETMamba: a pet maximum-intensity-projection-based artificial intelligence framework for aetiological classification of fever of unknown origin with multicentre validation.
Authors
Affiliations (8)
Affiliations (8)
- Department of Nuclear Medicine, West China Hospital, Sichuan University, Chengdu, Sichuan, China.
- West China Biomedical Big Data Center, West China Hospital, Sichuan University, Chengdu, China.
- Department of Nuclear Medicine, Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School, Nanjing, China.
- Department of Oncology, West China Hospital of Sichuan University, Chengdu, China.
- Department of Hematology, Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School, No. 321, Zhongshan Road, Nanjing, Jiangsu, 210008, China. [email protected].
- Department of Nuclear Medicine, The First Affiliated Hospital of Nanjing Medical University, Jiangsu Province Hospital, No. 321, Zhongshan Road, Nanjing, Jiangsu Province, 210008, China. [email protected].
- Shanghai Medical College, Fudan University, No. 130, Dongan Road Xuhui District, Shanghai, 200032, China. [email protected].
- Department of Nuclear Medicine, West China Hospital, Sichuan University, Chengdu, Sichuan, China. [email protected].
Abstract
Fever of unknown origin (FUO) remains diagnostically challenging because of heterogeneous causes, non-specific clinical manifestations, and overlapping imaging findings. We developed and validated FUO-PETMamba, a PET maximum-intensity-projection (MIP)-based artificial intelligence framework for AI-assisted aetiological classification of FUO. This retrospective multicentre study included 681 patients with FUO who underwent baseline [18 F]FDG PET/CT, comprising one development cohort (n = 355) and two independent external validation cohorts (n = 195 and n = 131). FUO-PETMamba is a weakly supervised framework analysing PET MIP images generated from PET data. Model performance was assessed by discrimination, calibration, decision curve analysis (DCA). Attention-based visual explanations were generated using attention mechanisms and gradient-based activation mapping. A reader study assessed the potential assistive effect of model-predicted probabilities on physicians with different PET/CT experience. In the development cohort, FUO-PETMamba achieved AUCs of 0.838 for malignancy, 0.851 for infection, 0.914 for autoimmune disease, and 0.788 for miscellaneous causes. Corresponding AUCs were 0.809, 0.815, 0.805, and 0.849 in external validation cohort 1, and 0.808, 0.772, 0.701, and 0.927 in external validation cohort 2, respectively. Calibration and decision curve analysis suggested potential clinical benefit for the major aetiological categories, although performance varied across cohorts and classes. The miscellaneous category should be interpreted cautiously because of limited case numbers and low positive predictive value and F1 scores. In the reader study, AI assistance improved diagnostic accuracy for junior and intermediate physicians, whereas changes in senior-physician performance were variable. Post hoc attention-based visualisations highlighted clinically plausible hypermetabolic patterns and served as qualitative explanatory aids. FUO-PETMamba provides a PET MIP-based AI-assisted diagnostic support framework for aetiological classification of FUO across multicentre cohorts and may help reduce experience-dependent diagnostic variability after prospective validation.