AI-Driven segmentation and morphogeometric profiling of epicardial adipose tissue in type 2 diabetes

Figures

• 

  Fig. 1. Overview of the epicardial adipose tissue (EAT) segmentation and analysis pipeline for investigating morphogeometric differences between type 2 diabetes (T2D) and control groups. The workflow begins with Dixon MRI acquisition, followed by automatic EAT segmentation using a deep learning-based method (EAT-Seg). Statistical shape analysis is performed on point cloud representations of EAT across all samples. Partial least squares-discriminant analysis (PLS-DA) is applied to capture spatial differences between groups through latent variables. Morphogeometric features, including geometric and shape-based metrics, are extracted and analyzed for their correlation with latent variables. Features showing high correlations are identified as key contributors to group separation. Random Forest model is used to evaluate these features for identifying T2D-specific EAT characteristics

• 

  Fig. 2. The Dixon MRI sequence consists of four phases: fat phase (FP), water phase (WP), out-of-phase (OP) and in-phase (IP) images. Each phase contains the same number of slices. Epicardial adipose tissue (EAT) is identified as interior pixels highlighted in orange. The different phases show varying contrasts that contain both structural and contextual information. 3D reconstruction of EAT segmentation from Dixon MRI is shown in the bottom row

• 

  Fig. 3. Overview of the EAT-Seg architecture and its components. A the model follows a U-shaped architecture with Res-Mamba blocks in the encoder and decoder paths, processing multi-phase Dixon MRI inputs to generate epicardial adipose tissue (EAT) segmentation masks and signed distance maps (SDMs). B the input encoder integrates multi-phase Dixon MRI data using an input-level fusion strategy and a channel attention mechanism to emphasize key features. C the Res-Mamba block enhances feature extraction by combining long-range dependencies (SSM-based Mamba block), multi-scale integration (channel fusion layer), and fine-grained feature preservation (Residual Block), improving segmentation in anatomically complex EAT regions

• 

  Fig. 4. Comparison of EAT-Seg with representative methods. Segmentation results for one T2D and one control subject are visualized in both 3D and 2D slices. The top row presents 3D reconstructions of the ground truth (orange) and surface distance error maps for different methods. The subsequent rows show axial slices at cranial, middle, and caudal regions of the heart, with segmentation overlays. The ground truth is shown in orange, while the predicted results are overlaid in violet. DSC, Dice similarity coefficient; T2D, type 2 diabetes

• 

  Fig. 5. Statistical shape analysis of epicardial adipose tissue (EAT) in type 2 diabetes (T2D) and control groups. A PLS-DA separates T2D and control groups based on EAT point cloud data, with Latent Variable 1 distinguishing between groups and Latent Variable 2 capturing intra-group variability. Representative samples from each group are visualized as 3D point clouds (top) and corresponding EAT reconstructions (bottom), highlighting structural differences in EAT morphology. B spatial distribution of high-contribution regions identified by variable importance in projection (VIP) scores. Regions with VIP scores exceeding 1 (top) and 3 (bottom) highlight key spatial differences between T2D and control groups

• 

  Fig. 6. Local thickness maps of epicardial adipose tissue (EAT) for representative samples from type 2 diabetes (T2D) and control groups. These six samples correspond to those shown in Fig. 5A. The color-coded thickness maps illustrate regional variations in EAT thickness. T2D samples exhibit relatively higher average thickness and greater spatial heterogeneity compared to controls