Automatic Segmentation of Epicardial Adipose Tissue Volume from cardiac HASTE MRI using 3D nnU-Net

Background: Epicardial adipose tissue (EAT) is a visceral fat surrounding the myocardium. Elevated EAT is an independent predictor of major adverse cardiovascular events (MACE), diabetes, and coronary artery disease (CAD) ^1-3. Manual segmentation of EAT from CMR is laborious. Recently, an automated approach for EAT segmentation was proposed⁴; however, the 3D Dixon imaging sequence utilized for model development has a long acquisition time and is not commonly utilized in a clinical setting. Our goal was to train a deep learning model with Half-Fourier Single-Shot Turbo Spin Echo (HASTE) imaging, a rapid clinical sequence routinely used for cardiac morphology evaluation that provides excellent contrast between fat and myocardium and delineation of the pericardium, making it potentially suitable for EAT quantification.

Methods: HASTE short-axis images were acquired on a 3T MRI scanner (MAGNETOM Vida, Siemens Healthineers) from healthy individuals and patients with heart failure with preserved ejection fraction (HFpEF). Parameters included: TR = 800 ms, TE = 33 ms, matrix = 192×256, in-plane resolution = 1.6×1.6 mm, and slice thickness = 6 mm. EAT was manually segmented in 3D Slicer from the cardiac base to apex. Datasets were split into training, validation, and test sets, with each subject assigned to a single split to avoid bias from repeated imaging. The self-configuring nnU-Net framework was used with the 3D full-resolution U-Net trained per the original implementation⁵. Five-fold cross-validation was performed, and the five models were ensembled for testing. Performance was assessed using Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (HD95). Concordance correlation coefficient (CCC) and linear regression was evaluated to assess agreement between predicted and reference EAT volumes.

Results: Fifteen healthy subjects and eighteen patients with HFpEF were included. A total of 59 examinations were used for model development (39 training; 7 validation; 13 test). The average DSC from the test set was 0.84 and HD95 was 8 mm. There was strong agreement between the model predicted and reference EAT volumes (CCC = 0.96). The model systematically underestimates the EAT volume, as seen in the regression plot (Figure 2). Figure 3 presents model predicted and reference EAT segmentations across several slices for a representative examination.

Conclusion: Using rapid single-shot HASTE images commonly acquired to assess cardiac morphology, we were able to achieve automated EAT segmentation results comparable to what was previously demonstrated using 3D Dixon images. Additionally, the self-configuring nnU-Net enabled a feasible training scheme. While the model underestimated the EAT volume, this may be addressed with additional data or data augmentation. Nonetheless, there was strong agreement between the volumes from both methods. Thus, this nnU-Net framework is a suitable tool for automatic EAT volume evaluation and may be extended to include other clinical imaging sequences.

Figure 1: nnU-Net framework utilized for model training and testing. Short-axis HASTE images are automatically pre-processed using the self-configuring pipeline to determine the optimal input matrix across the datasets. The resized images are input into a U-Net architecture with filter sizes of 32, 64, 128, 256, and 320 and a 3D full resolution training scheme.
Figure 2: Scatter plot of EAT volumes from reference and model predicted segmentations. The regression line is shown in red. There was strong agreement between the reference and model predicted EAT volumes (CCC=0.96). The regression line slope indicates that the model systematically underestimates the EAT volume.
Figure 3: Representative examination illustrating the model prediction and reference segmentations across slices of the cardiac base to apex. The model underestimates the EAT on the inferior surface of the heart and around the apex, as indicated with white arrows.