MorVess: Morphology-Aware Pulmonary Vessel Segmentation Network

Technical Summary: MorVess – Morphology-Aware Pulmonary Vessel Segmentation Network

Problem Statement
Accurate segmentation of pulmonary vessels remains a significant challenge in medical image analysis due to the sparse, tortuous, and multi-scale nature of vascular structures. Existing deep learning approaches primarily optimize binary masks using voxel-wise supervision, which often fails to preserve the continuous tubular morphology and fine connectivity of small branches. Traditional methods and standard deep learning models struggle with topological discontinuities, boundary ambiguity, and diameter inconsistencies, particularly in distal microvascular regions where vessels are slender and low-contrast. Furthermore, while foundation models like the Segment Anything Model (SAM) show promise, their direct application to 3D medical data is hindered by a lack of 3D spatial context (when processed slice-by-slice) and an absence of explicit geometric constraints necessary for tubular structures.

Methodology
The authors propose MorVess, a morphology-aware segmentation framework that integrates differentiable geometric priors with large-scale foundation model adaptation. The architecture consists of three core components:

1. Geometric Prior Generation: To address boundary ambiguity and diameter discontinuity, the method generates two continuous, differentiable supervisory signals from ground-truth masks:

   • Vessel Distance Map (VDM): Transforms binary mask boundaries into smooth gradient potential fields using morphological erosion and exponential distance decay. This provides sub-pixel localization and sharpens boundary gradients.
   • Vessel Thickness Map (VTM): Derived from the vascular centerline skeleton, this map assigns thickness values to voxels based on the maximal inscribed sphere radius. It enforces global consistency in vessel diameter and prevents non-physical oscillations in local regions.

2. 2.5D Geometric Adaptation Network:

   • 2.5D Adapter: To bridge the gap between 2D pre-trained SAM representations and 3D volumetric data, a lightweight adapter is introduced. It processes input volumes as sequences of overlapping slices (e.g., a target slice plus adjacent neighbors) and uses a 3D convolutional structure to model inter-slice dependencies without unfreezing the heavy SAM ViT encoder.
   • Multi-head Geometric Prediction Decoder: Extending the SAM decoder, this module includes parallel branches to jointly regress the semantic mask, VDM, and VTM.
   • Global-Local Fusion Block (GLFB): This block aggregates multi-level features (global semantic context, shallow local textures, and deep semantic abstractions) along with the self-predicted geometric priors (VDM/VTM). A dynamic hypernetwork mechanism adaptively weights these features to refine the final vascular topology.

3. Two-Stage Training Strategy:

   • Stage I (Macro-Feature Adaptation): The SAM encoder is frozen while the 2.5D adapter and decoder heads are trained to learn inter-slice relationships and map geometric priors.
   • Stage II (Topological Fine-tuning): The adapter is frozen, and the fusion module and prediction heads are fine-tuned with a lower learning rate to optimize boundary precision and topological consistency.

The training utilizes a composite loss function combining Cross-Entropy, Dice, centerline Dice (clDice), and L1 losses for the distance and thickness maps.

Key Contributions

• Geometry-Prior Segmentation: Introduction of a joint optimization framework for Vessel Thickness Maps (VTM) and Vessel Distance Maps (VDM), which explicitly reinforces the preservation of vascular topology and coherent geometric structures across complex branching patterns.
• Lightweight 2.5D Framework: Design of a cross-slice attention framework with an enhanced fusion block that integrates multi-level spatial cues, strengthening inter-slice coherence and improving the continuity of fine-scale vessels while maintaining computational efficiency.
• Resource-Efficient Adaptation: Development of a two-stage training strategy that enables stable adaptation of large pre-trained models (SAM) to 3D medical segmentation tasks under limited computational budgets.

Experimental Results
MorVess was evaluated on two challenging pulmonary CT benchmarks: Parse2022 (focused on arterial structures) and AIIB2023 (featuring pulmonary fibrosis with complex pathologies).

• Performance Metrics: MorVess achieved superior performance across all metrics compared to state-of-the-art methods (including nnU-Net, Swin-UNet, and COMMA).
  • On Parse2022, it achieved a Dice score of 86.84% (vs. 83.27% for the second-best) and a clDice of 83.22%.
  • On AIIB2023, it reached a Dice of 94.31% and clDice of 89.34%.
  • Significant improvements were also observed in HD95 (boundary accuracy), AMR (missing rate), DBR (branch detection ratio), and DLR (length detection ratio).
• Ablation Studies: Experiments confirmed that the inclusion of both VDM and VDM supervision synergistically improved performance, reducing HD95 from 6.85 mm to 4.53 mm on Parse2022. The 2.5D adapter and GLFB were also shown to be critical for 3D continuity and fine-detail reconstruction.
• Cross-Dataset Generalization: The model demonstrated robustness when applied to out-of-distribution datasets (HiPas and ATM2022) without fine-tuning, maintaining high Dice and clDice scores, indicating that the geometric constraints encode topological continuity rather than dataset-specific appearance.
• Geometric Analysis: Using Vascular Modeling Toolkit (VMTK), MorVess showed the lowest error in Total Vessel Volume (TVV) and highest correlation in diameter distribution compared to ground truth, validating its ability to preserve physiological vessel characteristics.
• Efficiency: The model requires only 1.0M trainable parameters and 42 GMACs per slice stack, significantly lower than competing methods like nnU-Net (32M params, 180 GMACs).

Significance and Claims
The paper claims that MorVess offers a principled and scalable pathway toward precise vessel analysis by embedding geometric intelligence into pre-trained vision models. By moving beyond simple binary mask optimization to include explicit, differentiable geometric constraints (distance and thickness), the framework addresses the fundamental limitations of current models in recovering continuous tubular morphology. The authors assert that this approach enables clinically reliable structural quantification, particularly for small-vessel recovery and global connectivity, which are critical for pulmonary disease diagnosis and preoperative planning. The work highlights that combining foundation model adaptability with morphology-aware supervision can overcome the trade-off between high accuracy and the computational cost of 3D processing.