Cross-Cohort Federated Learning for Pediatric Abdominal Adipose Tissue Segmentation and Quantification Using Free-Breathing 3D MRI

Figures

• 

  Figure 1. Cross-cohort federated training strategy and 3D neural network architecture. (A) Diagram illustrating the federated training strategy to train a 3D neural network for abdominal adipose tissue segmentation on 3D MRI. The two training datasets, adult and pediatric MRI, were maintained in distinct local environments without exchanging data at any point. Only the encoder was shared, and the model adaptor and decoder were retained locally to preserve domain-specific data characteristics. (B) Detailed depiction of the processing pipeline with multiple MRI series input into local 3D neural networks composed of a shared encoder and task-specific local decoders. The neural network backbone was based on 3D nnU-Net. The definitions of proton-density fat fraction (PDFF) and signal fat fraction (SFF) are provided in the Supplemental Materials.

• 

  Figure 2. Example pediatric abdominal adipose tissue segmentation results using the federated learning model in a 12-year-old female. This was a lean subject with body mass index of 20.43 kg/m2 (71.8 percentile) who exhibited the median Dice scores (93.13% for VAT and 98.48% for SAT) using the neural network model in the testing set. (A) Input MRI series, reference label, and model segmentation results. The same display window/level was used for the TEOP, TEIP, water, and fat images. Grayscale bars (0%-100%) are shown for proton-density fat fraction (PDFF) maps. (B) False negative and false positive voxels. The false negative rate across all SAT/VAT voxels was 0.38%, and the false positive rate was 0.08%. Definitions and equations are provided in the Supplemental Materials.

• 

  Figure 3. Example pediatric abdominal adipose tissue segmentation results using the federated learning model in a 13-year-old female. This was a very lean subject with body mass index of 15.64 kg/m2 (6.5th percentile) who exhibited the lowest Dice scores (78.38% for VAT and 77.00% for SAT) using the neural network model in the testing set. (A) Input MRI series, reference label, and model segmentation results. The same display window/level was used for the TEOP, TEIP, water, and fat images. Grayscale bars (0%-100%) are shown for proton-density fat fraction (PDFF) maps. (B) False negative and false positive voxels. The false negative rate across all SAT/VAT voxels was 1.02%, and the false positive rate was 0.01%. Definitions and equations are provided in the Supplemental Materials.

• 

  Figure 4. Model performance results for abdominal adipose tissue segmentation using MRI. (A) Box plots summarizing the segmentation Dice scores in the pediatric MRI testing set using the models obtained from standalone (P1), federated (FL1), and joint (JT) training. Results from individual subjects are also plotted as circles. Mean Dice scores for SAT segmentation were: 93.7% (P1), 95.55% (FL1), and 95.94% (JT). For VAT segmentation, mean Dice scores were 87.57% (P1), 91.09% (FL1), and 90.69% (JT). (B) Plots showing the regression analysis of segmentation Dice scores with respect to body mass index (BMI) percentile in the pediatric testing set. Both SAT and VAT segmentation performance increased as the BMI percentile increased. (C) Summary of Dice scores for the adult MRI testing set. (D) Regression analysis of segmentation Dice scores with respect to BMI in the adult testing set. See Tables 3 and 4 for more details. Following the Kruskal–Wallis H-test, we found significant differences in VAT segmentation performance across the 3 models (corrected P = .03, η2=0.08), while no significant differences were detected for SAT (corrected P = .74, η2=0). Following the pairwise 1-sided Wilcoxon…

• 

  Figure 5. Abdominal VAT and SAT volume quantification results. (A) Abdominal VAT volume quantification in the pediatric MRI testing set using the standalone (P1), federated (FL1), and jointly trained (JT) model segmentation results compared to reference manual measurements. The intraclass correlation coefficients (ICCs) are reported. (B) Abdominal SAT volume quantification in the pediatric MRI testing set using the P1, FL1, and JT model segmentation results compared to reference manual measurements. The ICCs were all close to unity. (C, D) Abdominal VAT and SAT volume quantification results in the adult MRI testing set. All models achieved ICC close to unity with respect to references.

• 

  Figure 6. Pediatric abdominal VAT and SAT proton-density fat fraction (PDFF) quantification results. (A) Bland-Altman analysis of VAT PDFF calculated using segmentation results of the standalone local model, the federated learning (FL) model, and the jointly trained model compared to reference manual measurements in the testing set, presented separately for each model. (B) Bland-Altman analysis of SAT PDFF calculated using segmentation results of the standalone local model, the federated learning model, and the jointly trained model compared to reference manual measurements in the testing set, presented separately for each model. For joint training and FL, the mean differences (MD) were small (0.21% to −1.39%), and the 95% limits of agreement (LoA) were less than ±10%.