Cross-cohort federated learning for pediatric abdominal adipose tissue segmentation and quantification using
Wenwen Zhang1,2, Sevgi Gokce Kafali1,3, Timothy Adamos4
1Department of Radiological Sciences, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, 90095, United States.
Background:
Pediatric abdominal visceral and subcutaneous adipose tissue (VAT, SAT) quantified on magnetic resonance imaging (MRI) can assess risk for metabolic diseases. However, the complex structure of VAT in children and the lack of sufficient MRI datasets pose challenges for developing automated segmentation methods.
Purpose:
To achieve accurate and rapid automated segmentation of pediatric abdominal VAT and SAT on motion-robust free-breathing (FB) 3D Dixon MRI by developing a cross-cohort federated learning (FL) framework that leverages adult datasets.
Materials And Methods:
3D FB-MRI datasets were prospectively acquired in children 6-18 years old (single center, 2 scanners; 2016-2023) and used to train 3D neural network models for segmenting abdominal VAT and SAT. The FL model was trained across the pediatric cohort and a separate adult cohort (5 centers, 7 scanners) without requiring direct data sharing. Segmentation performance of the FL model was assessed by Dice scores with respect to references and compared with standalone local training and joint training with full data access. Quantification of VAT and SAT volume and proton-density fat fraction (PDFF) was compared against references using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. Differences between training approaches were analyzed using the Kruskal-Wallis test followed by paired Wilcoxon signed-rank tests.
Results:
The FL model, trained and tested with 134 children (mean age, 13.3 years ± 2.7 [standard deviation]; 71 males) and 920 adults (50.4 years ± 14.0; 677 females), achieved mean Dice scores of 91.09% (VAT) and 95.55% (SAT), outperforming standalone training (VAT: P < .001) and performing comparably to joint training (VAT: P = .21). Volume quantification demonstrated strong agreement (VAT: ICC = 0.99, SAT: ICC = 1.00). PDFF quantification showed small mean differences (VAT: 0.21%, SAT: -1.19%). Inference time was <3 seconds for each subject.
Conclusion:
The proposed FL framework achieved accurate and rapid automated segmentation and quantification of pediatric abdominal VAT and SAT on 3D FB-MRI.
