This week, Team Genesis focused on improving the smoothness and quality of our liver meshes for simulation. We experimented with higher-resolution meshing, introduced new smoothing techniques, and investigated methods to detect and repair mesh defects.

We also discussed our progress with Dr. Catia Silva, focusing on the trade-offs between mesh quality, anatomical accuracy, and physics readiness.

Improving Liver Mesh Smoothness

One of our main experiments involved improving the surface quality of meshes generated using fTetWild. Our initial meshes contained visible triangular patterns and uneven surfaces that could affect their suitability for simulation.

We compared three approaches: original fTetWild, high-resolution fTetWild, and Signed Distance Field (SDF) smoothing combined with fTetWild.

Increasing the resolution improved surface detail, while SDF smoothing helped reduce visible irregularities. We also explored Taubin smoothing to further improve mesh quality while limiting unwanted changes to the liver’s shape.

Comparing Different Meshing Approaches

Our experiments showed clear differences in surface quality and mesh complexity. The original fTetWild mesh contained approximately 12,582 boundary faces, while the high-resolution version increased this to 166,172 faces.

Combining SDF smoothing with fTetWild produced a smoother surface using approximately 69,506 boundary faces, demonstrating a promising balance between mesh complexity and visual quality.

These results suggest that increasing triangle counts alone may not be necessary to achieve smoother anatomical models.

Strengthening Mesh Quality and Repair

Another important focus was identifying mesh defects that could prevent our models from being used in physics simulations.

We investigated self-intersecting triangles, non-manifold edges, and poor triangle quality. We also explored repair techniques such as MeshFix and targeted hole filling to address issues that padding alone could not resolve.

To track these problems systematically, we continued developing our quality-control framework and proposed a failure mode register to document issues across the pipeline.

Looking Ahead

Next, we will compare mesh quality across different triangle counts, investigate localized smoothing techniques, and evaluate how effectively automated repair methods resolve remaining mesh defects.

We also plan to begin testing our reconstructed liver assets in NVIDIA Isaac Sim to evaluate their physical behavior and move closer to producing reliable, physics-ready anatomical models.

This week, Team Genesis focused on making our liver models more reliable for simulation. We explored different mesh generation techniques, studied how small changes to segmentation masks affect mesh quality, and developed new quality-control checks for our pipeline.

We also completed our Preliminary Design Review (PDR) presentation with our NVIDIA liaison, Oge Marques, discussing our progress and the challenges ahead.

Improving Liver Mesh Reliability

One of our main experiments involved testing 102 liver segmentation masks from TotalSegmentator. We compared different levels of zero padding before converting the masks into 3D meshes.

Adding padding improved the percentage of watertight meshes from 63% to 83%. However, we also discovered that excessive padding can change the liver’s anatomical shape, making it less suitable for applications such as robotic ultrasound training.

To address this, we proposed limiting how much extra volume padding can introduce, ensuring that models used for robot training remain anatomically accurate.

Comparing Different Meshing Approaches

We also compared marching cubes with fTetWild, a tetrahedral meshing approach. While marching cubes produced smoother surfaces, fTetWild generated watertight meshes with fewer vertices in our initial comparison.

These results are promising, but further testing across additional liver samples is needed before selecting the most suitable approach.

Building a Quality-Control Framework

Another important development was creating a quality-control framework to evaluate our pipeline at multiple stages, from image ingestion and segmentation to mesh validation and physics readiness.

We tested this framework using three liver datasets: TotalSegmentator, FLARE22, and MSD Liver. We also developed structured reports that record whether each stage passes or fails and explain the reasons for failure.

Our next improvements will focus on detecting mesh self-intersections, evaluating triangle quality, and checking whether the reconstructed models preserve the original anatomy.

Looking Ahead

Next, we will expand our meshing experiments, refine the quality-control thresholds, and continue improving the reliability of liver models for simulation. We will also use feedback from our NVIDIA liaison to guide the next steps toward producing physics-ready anatomical assets.

Week 3 marked an important step forward for Team 17 as we moved from initial research toward defining our technical approach. Throughout the week, we explored the major stages of our pipeline, including CT image segmentation, 3D mesh generation, and quality control, while comparing different approaches with Dr. Silva. We also identified several potential public CT datasets and began narrowing down the options based on the anatomy that will best support our project. As we prepare for our upcoming meeting with the NVIDIA liaison engineer, the team has compiled key technical and scope questions to clarify project expectations and ensure that our approach aligns with NVIDIA’s goals. With our initial research and task assignments complete, we are now looking forward to setting up NVIDIA Isaac Sim on HiPerGator, investigating TotalSegmentator, and taking our first steps toward building the pipeline