KidMesh: Computational Mesh Reconstruction for Pediatric Congenital Hydronephrosis Using Deep Neural Networks

1. Problem Statement

Pediatric Congenital Hydronephrosis (CH) is a common urinary tract disorder characterized by the swelling of the renal pelvis and calyces due to obstruction. While Magnetic Resonance Urography (MRU) provides excellent anatomical visualization, current clinical workflows face significant limitations:

• Voxel-based Limitations: Traditional segmentation methods (e.g., U-Net) output voxel grids. Converting these to 3D meshes for Computational Fluid Dynamics (CFD) simulations requires complex post-processing (e.g., Marching Cubes, smoothing, topology correction).
• Artifacts and Quality: Due to sparse sampling in MRU (especially along the Z-axis), voxel-based reconstructions often suffer from "staircase artifacts," leading to poor geometric quality and topological errors (e.g., non-watertight surfaces).
• Lack of Functional Data: Standard imaging provides structural data but lacks the urodynamic (flow/pressure) insights necessary for assessing disease severity and surgical outcomes.
• Annotation Scarcity: High-quality, vertex-level mesh annotations for training deep learning models are extremely difficult to obtain manually, limiting the adoption of supervised mesh reconstruction methods.

2. Methodology: KidMesh Framework

The authors propose KidMesh, an end-to-end deep learning framework that directly reconstructs simulation-ready 3D meshes from MRU images without requiring manual mesh-level annotations. The architecture consists of three core modules:

A. Feature Extraction Module (FEM)

• Architecture: A 3D U-Net-like encoder-decoder network.
• Function: Extracts hierarchical feature maps from the input MRU volume.
• Strategy: Uses deep supervision to generate binary segmentation masks, ensuring the network focuses on boundary-sensitive features critical for identifying the hydronephrosis region.

B. Feature Sampling Module (FSM)

• Function: Acts as the bridge between volumetric features and mesh vertices.
• Mechanism:
  • Grid Sampling: Projects 3D feature maps onto mesh vertices using trilinear interpolation.
  • Adaptive Neighborhood: Learns offset vectors to define local sampling neighborhoods around each vertex.
  • Graph Convolution & Self-Attention: Aggregates local features via Graph Convolutional Networks (GCN) and employs a parameter-free self-attention mechanism to model long-range dependencies between anatomically related vertices, ensuring global geometric coherence.

C. Mesh Deformation Module (MDM)

• Function: Deforms a spherical template mesh into the target CH shape.
• Strategy: A coarse-to-fine progressive deformation process.
  • Dynamic Vertex Upsampling (VU): Uses Uniform Unpooling (UU) to subdivide mesh faces, increasing resolution.
  • Vertex Filter (VF): A novel mechanism that prunes vertices with "no contribution" (stationary) or "negative contribution" (excessive displacement) to maintain geometric stability and prevent self-intersections.
• Output: Generates a watertight, topologically consistent mesh suitable for CFD.

D. Training Strategy: Weak Supervision

• Challenge: Lack of ground-truth mesh annotations.
• Solution: The model is trained using pseudo-gold standard meshes generated automatically from expert-annotated voxel masks via the Marching Cubes algorithm. These pseudo-meshes retain staircase artifacts and noise.
• Loss Function: A hybrid loss function guides the network:
  • Extraction Loss: Dice + Cross-Entropy (for the segmentation branch).
  • Reconstruction Loss: Chamfer Distance (geometric similarity) + Regularization terms (Laplacian, Normal, Edge Length, Area, and a novel Seal Loss).
  • Seal Loss: A differentiable term designed to penalize surface validity failures (e.g., face collapse, normal inversion) to ensure watertightness without discrete topological checks.

3. Key Contributions

1. First End-to-End Framework for Pediatric CH: KidMesh is the first method to generate CFD-ready meshes directly from MRU scans for pediatric hydronephrosis, eliminating the need for manual post-processing.
2. Dynamic Vertex Upsampling (VU): Introduces a coarse-to-fine pipeline with adaptive refinement (UU) and filtering (VF) to handle the heterogeneous geometry of renal calyces while maintaining stability.
3. Weakly Supervised Learning: Successfully trains a mesh reconstruction network using noisy, artifact-laden pseudo-meshes derived from voxel masks, removing the bottleneck of manual mesh annotation.
4. CFD-Ready Output: The method guarantees topological correctness (watertightness, no self-intersections), enabling direct use in fluid dynamics simulations.

4. Experimental Results

The study was conducted on MRU data from 160 pediatric patients (112 training, 32 validation, 16 testing).

• Performance Metrics:

  • Speed: KidMesh reconstructs a mesh in an average of 0.36 seconds, compared to ~86.5 seconds for voxel-based methods (RES50 + post-processing).
  • Accuracy: Achieved a Dice score of 0.86 (after rasterization) against manual masks.
  • Geometric Quality: Only 3.7% of vertices had error distances >3.2mm, and 0.2% exceeded 6.4mm. The reconstructed meshes exhibited no self-intersections.
  • Comparison: Outperformed state-of-the-art explicit methods (MeshDeformNet) and voxel-based baselines (nnU-Net, FCN) in terms of Hausdorff Distance (HD) and Point-to-Surface Distance (P2SD), while producing smoother surfaces free of staircase artifacts.

• Surface Validity: KidMesh significantly reduced topological failures (dangling nodes, self-intersecting polygons) compared to MeshDeformNet, demonstrating superior robustness.

• Urodynamic Simulation (CFD):

  • Simulations performed on KidMesh meshes showed high quantitative agreement with simulations on manual reference meshes (mean velocity difference <3.6%, pressure discrepancy <1.0 Pa).
  • The model successfully captured flow acceleration at the ureteropelvic junction and pressure gradients, validating its utility for functional analysis.
  • Pre- vs. post-operative simulations demonstrated the ability to visualize changes in wall shear stress and pressure, offering potential clinical insights into surgical efficacy.

5. Significance and Impact

• Clinical Workflow Integration: By automating the transition from MRU to simulation-ready mesh, KidMesh drastically reduces the time and expertise required for patient-specific urodynamic analysis.
• Functional Assessment: It bridges the gap between structural imaging and functional assessment, allowing clinicians to visualize urine flow dynamics and pressure changes, which are critical for diagnosing obstruction severity and planning interventions.
• Generalizability: The weakly supervised approach and topology-preserving design make the framework applicable to other fluid-filled anatomical structures where high-quality mesh annotations are scarce.
• Future Directions: The authors suggest extending the framework to model the entire urinary tract (ureters and bladder) and improving performance in high-curvature regions through adaptive mesh refinement.

In conclusion, KidMesh represents a significant advancement in medical image analysis, transforming static MRU data into dynamic, simulation-ready 3D models that can directly support clinical decision-making for pediatric congenital hydronephrosis.