Deep-Learning Quantification of Epicardial Adipose Tissue based on 3D Dixon MRI

Presenting Author(s)

• 

  Halil Noyan, MD

  Medical Student
  Charité – Universitätsmedizin Berlin
  Berlin, Berlin, Germany

Primary Author(s)

• 

  Halil Noyan, MD

  Medical Student
  Charité – Universitätsmedizin Berlin
  Berlin, Berlin, Germany

Co-Author(s)

• 

  Richard Hickstein, MD

  Physician and Researcher
  Charité – Universitätsmedizin Berlin
  Berlin, Berlin, Germany
• CL

  Christian Lim

  Clinical Researcher
  Working Group on CMR, Experimental and Clinical Research Center, a cooperation between Charité – Universitätsmedizin Berlin and the Max Delbrück Center for Molecular Medicine in the Helmholtz Association, Berlin, Germany, Germany
• 

  Clemens Ammann, MD

  Physician
  Charité – Universitätsmedizin Berlin
  Berlin, Berlin, Germany
• 

  Thomas C. R Hadler, PhD

  Postgraduate
  Charité - Universitätsmedizin Berlin
  Berlin, Berlin, Germany
• CH

  Christian Hickstein

  Researcher
  Technische Universität Berlin, Faculty IV – Electrical Engineering and Computer Science, Institute of Software Engineering and Theoretical Computer Science, Straße des 17. Juni 135, 10623 Berlin, Germany, Germany
• JK

  Johanna Kuhnt

  Medical student
  Charité – Universitätsmedizin Berlin, Germany
• ED

  Elias Daud, MD

  Physician
  Charité – Universitätsmedizin Berlin, Germany
• 

  Edyta Blaszczyk, MD

  Cardiologist
  Charité - Universitätsmedizin Berlin
  Berlin, Berlin, Germany
• 

  Jeanette Schulz-Menger, MD

  Head Working Group Cardiac MRI
  Charité/ University Medicine Berlin and Helios
  Berlin, Berlin, Germany

Background: Epicardial adipose tissue (EAT) is increasingly being recognized as a relevant biomarker linked to coronary artery disease, atrial fibrillation, and adverse cardiovascular outcomes (1,2). High-resolution isotropic (1.3 mm³) 3D Dixon water–fat imaging enables detailed visualization and quantification of epicardial fat (3). However, manual EAT quantification is time-consuming and reader-dependent, limiting its integration into clinical routine.

We aimed to develop and evaluate a deep learning–based approach for automated ventricular EAT segmentation from 3D Dixon data.

Methods: We implemented a 3D full-resolution no-new-U-Net (nnU-Net) model on a Linux workstation for automated segmentation of ventricular EAT (4). Ground-truth (GT) segmentation was carried out in 3D Slicer (5). The training dataset involved 44 cases (36 with, 8 without contrast agent) with segmentations from two readers (O₁ and O₂). For benchmarking, nnU-Net segmentations were compared with GT segmentations in ten unseen cases. We evaluated the model using the Dice Similarity Coefficient (DSC), the 95th percentile Hausdorff distance (HD95), the absolute volume difference (ΔVol), the relative volume difference (RVD), Pearson correlation coefficients, intraclass correlation coefficients (ICC), and Bland–Altman analysis. To contextualize the results, intraobserver (O₁ vs. O₁) and interobserver (O₁ vs. O₂) variation was assessed in ten additional cases.

Results: The nnU-Net model vs. O₁ yielded a DSC of 0.80. Fig. 1 shows an example of the automated nnU-Net segmentation output, visualized in 3D Slicer, illustrating a close match between the segmented volume and the anatomical structures. The mean DSCs were 0.69 for the interobserver and 0.82 for the intraobserver agreement. A full overview of DSC, HD95, ΔVol, and RVD, including 95% confidence intervals, is provided in Tab. 1. Pearson correlation and intraclass correlation coefficients (ICC) showed excellent agreement for both O₁ vs. O₂ (r=0.98; ICC=0.90) and nnU-Net vs. O₁ (r=0.98; ICC=0.93). Corresponding scatter and Bland–Altman plots illustrate these agreements in Fig. 2. Segmentation time was reduced from approximately one hour for manual segmentation to less than ten seconds using the model.

Conclusion: A 3D nnU-Net–based approach enables fast ventricular EAT segmentation from 3D Dixon MRI, reducing processing time from nearly one hour to under ten seconds. The nnU-Net produces acceptable EAT volume estimations within the range of interobserver differences. It may support integration of EAT quantification into larger studies by overcoming the time constraints of manual processing.

Table 1: Quantitative comparison of segmentation performance. The Dice similarity coefficient (Dice), the 95th percentile Hausdorff Distance (HD95), the relative volume difference (RVD), and the absolute volume difference (ΔVol) are reported as mean ± standard deviation together with their 95 % confidence intervals (CI) in parentheses.

Comparison	Dice [-] (95% CI)	HD95 [mm](95% CI)	RVD [%] (95% CI)	ΔVol [ml] (95% CI)
nnU-Net modelvs. O1	0.80 ± 0.03(0.78–0.81)	2.01 ± 1.84(1.77–2.25)	8.55 ± 10.40(2.27–14.49)	5.81 ± 6.86(1.76–9.77)
O1 vs. O2	0.69 ± 0.12(0.62–0.76)	3.96 ± 2.35(2.65–5.41)	23.58 ± 20.86(12.92–37.03)	14.12 ± 7.93(9.59–18.86)
Intraobserver (O1)	0.82 ± 0.06(0.79–0.86)	1.70 ± 0.72(1.24–2.07)	0.94 ± 8.63 (-3.75–6.32)	-0.40 ± 4.77(-3.17–2.39)

Figure 1: Visualization of the fat phase from a 3D Dixon acquisition (top row) and the corresponding automated segmentation (yellow mask, bottom row) in 3D Slicer. Yellow arrows indicate epicardial adipose tissue. The red viewer shows a four-chamber view, the yellow viewer a two-chamber view, and the green viewer a short-axis view. The segmentation model correctly spares the coronary arteries (pink arrow). In the bottom row, the blue viewer presents a 3D reconstruction of the segmentation.
Figure 2: (a) Interobserver comparison between two readers (O1 vs. O2), (b) comparison between the nnU-Net model and reader O1, and (c) intraobserver comparison for reader O1. The top row shows Pearson correlation plots, and the bottom row shows the corresponding Bland–Altman plots. The limits of agreement in the Bland–Altman plots were defined as the mean difference ± 1.96 × the standard deviation of the differences.