Abstract: The preoperative planning of liver surgery relies on Couinaud segmentation from computed tomography (CT) images, to reduce the risk of bleeding and guide the resection procedure. Using 3D point-based representations, rather than voxelizing the CT volume, has the benefit of preserving the physical resolution of the CT. However, point-based representations need prior knowledge of the liver vessel structure, which is time consuming to acquire. Here, we propose a point-based method for Couinaud segmentation, without explicitly providing the prior liver vessel structure. To allow the model to learn this anatomical liver vessel structure, we add a graph reasoning module on top of the point features. This adds implicit anatomical information to the model, by learning affinities across point neighborhoods. Our method is competitive on the MSD and LiTS public datasets in Dice coefficient and average surface distance scores compared to four pioneering point-based methods.
The Challenge of Couinaud Segmentation
In liver surgery planning, surgeons rely on Couinaud segmentation, which divides the liver into eight functionally independent segments (I through VIII) based on the distribution of portal and hepatic veins. Accurate segmentation is vital to avoid major bleeding during resections, as each segment has its own vascular inflow, outflow, and biliary drainage.
Current automated methods face a "Goldilocks" problem. Voxel-based methods (CNNs) require resizing 3D CT volumes into fixed grids, which often destroys the physical resolution and fine details of the anatomy. Conversely, existing point-based methods preserve resolution by treating voxels as points in space, but they typically require "priors"—manual landmarks or pre-segmented vessel structures—to understand where one segment ends and another begins. This paper introduces a method to achieve high-resolution, point-based segmentation without requiring these labor-intensive manual vessel priors.
Methodology: Implicit Anatomical Learning
The researchers built their model, dubbed GrPn, on the Adaptive Graph CNN (AGCNN) architecture and introduced two key components to capture anatomical context without explicit labels:
- Grid Feature Embeddings: To provide spatial context, the model "revoxelizes" point features into local grids. This allows the use of 3D convolutions to extract local structural information before converting them back into point representations.
- 3D Dynamic Graph Reasoning: This module calculates "affinities" (relationships) between different neighborhoods in the liver. By learning these affinities, the model essentially "discovers" the underlying vessel structure implicitly, using it as a roadmap to define segment boundaries.
To manage the computational load of 3D data, the model uses a deformable unfold layer. This layer intelligently selects a small subset of informative voxels (a $3 \times 3 \times 3$ neighborhood) to compute affinities, significantly reducing memory overhead compared to global point-to-point calculations.
Experimental Results
The model was evaluated on two major public benchmarks: the Medical Segmentation Decathlon (MSD) and the Liver Tumor Segmentation Challenge (LiTS). Since these datasets do not natively include Couinaud labels, the authors utilized expert annotations to train and test their model.
GrPn outperformed several pioneering point-based baselines, including PointNet++ and the original AGCNN:
- Accuracy: It achieved the highest average Dice coefficient (a measure of overlap) on both datasets: 80.41% on MSD and 79.56% on LiTS.
- Boundary Precision: It achieved the lowest Average Surface Distance (ASD) on LiTS (5.50 mm), indicating that the predicted boundaries are physically closer to the true anatomical divisions.
- Qualitative Success: Visualizations show that the model's predicted boundaries align closely with the centerlines of the hepatic and portal veins, even though those veins were not explicitly labeled for the model.
Conclusion
The most significant finding is that deep learning models do not necessarily need to "see" a vessel to know it is there. By staying in the "point domain" and using graph reasoning to bridge local features and global structure, GrPn maintains the original spacing of medical images while reconstructing the anatomical logic of the segments. While the model currently faces challenges with large tumors that distort internal geometry, it represents a major step toward automated, high-resolution surgical planning without the need for manual preprocessing.
Summary updated