• 2019-06-17

Collaborators Meeting

Date: June 17-18, 2019

Means: In person at the University of Denver

Program:

Monday - June 17, 2019

Tuesday - June 18, 2019

Useful Links:

• Project website - https://simtk.org/projects/kneehub

• Project wiki - https://simtk.org/plugins/moinmoin/kneehub/

• Instructions for Model Development phase - https://simtk.org/plugins/moinmoin/kneehub/ModelDevelopment

• Data for Model Development phase - See ‘Data for MD - oks003’ & ‘Data for MD - DU02’ in the downloads section of the project site, https://simtk.org/frs/?group_id=1061

• Model Development Specifications - See the documents section of the project site, https://simtk.org/docman/?group_id=1061; also as submitted in source code repository https://simtk.org/svn/kneehub/ModelDevelopment/Specifications/

• Model Development Outcomes - as submitted in source code repository https://simtk.org/svn/kneehub/ModelDevelopment/Outcomes/

• Instructions for Model Calibration phase - https://simtk.org/plugins/moinmoin/kneehub/ModelCalibration

• Data for Model Calibration phase - See ‘Data for MC - oks003’ & ‘Data for MC - DU02’ in the downloads section of the project site, https://simtk.org/frs/?group_id=1061

Attendees:

1. Neda Abdollahi (CSU)
2. Thor Besier (ABI)
3. Shady Elmasry (HSS)
4. Ahmet Erdemir (CC)
5. Jason Halloran (CSU)
6. Donald Hume (DU)
7. Carl Imhauser (HSS)
8. Peter Laz (DU)
9. Nynke Rooks (ABI) - partially through Skype
10. Kevin Shelburne (DU)

Action Items:

• Ahmet
  • Evaluate the possibility of supplementary grant application to bring in outside groups to the project.
  • Arrange meeting in November 2019 to discuss planning and timeline of Model Benchmarking phase.
  • Arrange meeting midway through Model Calibration phase to evaluate processed calibration data of each team.
  • Update timeline of the project to reflect delays.
• Carl
  • Return back to JOR manuscript on reproducibility workshop to finalize and submit.

• Carl & Shady

  • Submit Jason review of knee modeling studies for reproducibility.
• Jason
  • Return back to manuscript on reproducibility potential of knee modeling studies to finalize data analysis and draft.
• Don
  • Share presentation on comparison of Model Development specifications.
• Thor
  • Outline publication on Model Development phase experience and outcomes.

• Thor & Nynke

  • Submit Model Development outcomes for oks003.
• All
  • Update presentation on comparison of Model Development specifications to accurately reflect individual team's workflow.
  • Prepare Model Development Specifications delivered, incorporating a-priori specifications and protocol deviations.
  • Prepare Model Calibration specifications for delivery by the end of July 2019.

Notes:

1. Model Development Phase Overview.

   • University of Denver started with a gross overview of their workflow. Don presented. Ligament insertions were determined by a combination of digitization data and segmentation (to confirm insertion points). They wanted to mark a slightly larger region of acceptability to lead into calibration, i.e., capturing plausible footprint. This was not a deviation, it noted in the specifications. Process was difficult for the collateral ligaments; they were thin and it was hard to see where they insert. Patellar mechanism was changed from the specifications, it was mostly an addition. DU02 mode was already calibrated in a previous publication but it will be further calibrated, the team has started from scratch. With oks003, they were having problem with patella bouncing of the knee; one of the collateral seems to be too tight. They also anticipate to incorporate meniscus contribution as a bulk during calibration. With the models Don played with ligament properties to make sure all contributed a bit to joint mechanics, essentially he performed a quick and dirty calibration. Abaqus Explicit was used and passive flexion was applied in a few seconds of simulation time. Simulations took about an hour. The membrane elements for the quad were driving the simulation time. In terms of effort and burden; as they have done similar modeling activities before their anticipated effort was pretty close, e.g. segment and build the models. All the documentation took longer. Prior experience helped with segmentation, model building, simulation. Quick notes included comments on cartilage meshing, morphing; no wrapping at tibiofemoral but at patellofemoral joint; patella kept as is in the image (slack). Don asked about quadriceps mechanism. Ahmet responded that it will come up in the reuse but not at calibration and benchmarking.

   • Cleveland Clinic shared their process next. Ahmet provided an overview of their workflow for building the models using Ariel's slides. They streamlined assembly of models from tissue surface representations where a connectivity file captures interactions between structures. First a template model was generated as a basic model (tetrahedral meshes, rigid bodies, placeholder materials), and then the template model was customized. A total of 16 tissue structures were segmented, all using 3D slicer. There was an emphasis on using free and open source software. Python and Salome were used to create 3D volume meshes of all the structures (all tetrahedral, except bones that are triangulated surfaces). Automated identification of interaction regions: e.g. region of cartilage contact to the bone were performed and corrected manually when needed. Anatomical landmarks were based on Grood and Suntay landmarks; due to lack of hip, long axis (z-axis of images) were used. The material representations in the customized model were more complex and inline with the specifications: bone = rigid, cartilage = Neo-Hookean, ligaments/tendon/meniscus = transversely isotropic Mooney-Rivlin. Pena paper was used for ligament properties. The team created a basic running model in FEBio with fully deformable soft tissues (bones rigid) and wrapping of all structures. Ahmet noted some nuances. FEBio doesn’t allow prescribing rigid body rotations for only in 1 axis (e.g. 2 free); it requires prescribing rotation around all 3 axes. Cylindrical joints were used to circumvent this problem (also noted in specifications). FEBio is not as mature and the team discovered a few bugs and documentation errors; these were reported to the developers to improve the software. Extensor mechanism had a protocol deviation; rather than an actuator to tension the quadriceps, they use a spring. Runtime was around 6 hours for oks003 and 1 hour for DU02, relying on implicit static analysis. An important deviation was the use of submodel (e.g., femur, tibia and ACL) to troubleshoot connectivity, which was really helpful to assess convergence and quality of model components, e.g. proofing ligament by ligament. Getting models to run was a significant effort (4 months, part time). There was a concern about ligament volumes, which leveraged input from a radiologist. Ahmet noted that these relied on interpretation during segmentation and smoothing. In response to a question, Ahmet confirmed that cartilage meshed was based on linear tetrahetrals and the team intend to consider quadratic tetrahedrals, which will be revisited as part of the calibration phase. He also mentioned that template modeling is a generalized workflow and it is agnostic to simulation software, e.g. models can be exported to different platforms. Ahmet also confirmed that all tissue structures were developed using MR data and they did not use probed point information. It will be interesting to see the comparisons in alignment between models.

   • Cleveland State University workflow was a hybrid between University of Denver and Cleveland Clinic workflows given Jason's experiences. His team leveraged the project as an opportunity on how to represent ligament anatomy, e.g. for structures that are less visible and to leverage literature. Jason used his team's protocol deviations document for presenting their work. Some deviations were data set specific, e.g., related to MRIs. Transfer of their experience in segmentation was a bit different due to differences in MRI of DU02 and oks003. Abaqus Explicit was used with a few seconds of simulation time; damping was important and it was selected based on prior knowledge, visual inspections, and as small damping and mass as possible. A large focus was on the anatomy of ligaments. Ligaments were represented as a series of interconnected springs (non-linear). This allowed wrapping around the bone. A separate simulation was performed to get wrapping of ligaments; overclosed ligaments were let “explode” and then snap back around the bone. This was done for all ligaments in one “explosion” step. This analysis didn't take much time. The team used Grow Cut segmentation in Slicer to improve segmentation efficiency and then proofed the segmented volumes. It will be interesting to compare process with Cleveland Clinic group. Meniscus and cartilage were meshed with hexahedral elements IA-FEMesh. Will made this work; the approach was similar to TrueGrid. Tetrahedrals were used for rigid bone. Non-linear springs represented horn attachments to connect menisci to bone. A basic set of Grood and Suntay coordinate system was used. Jason also shared an animation of a running model. Quadriceps representation was based on Open Knee(s) knowledge. Primarily the main quadriceps tendon and some retinaculum were represented. Simulation of passive flexion took about 2 hours. There was a mention of volume loss after smoothing. It would be interesting to compare raw and smooth geometries.

   • Thor described Auckland Bioengineering Institute work. Simulation using DU02 took 2 hours for full knee flexion. Model for oks003 has not be done yet and will be provided as soon as possible. Specifications for model development were inspired from for initial geometry and meshing workflow done by Mousa Kazemi, a prior PhD student but changed dramatically. The team thought that the major burden would be segmentation and mesh generation. They spent quite a time on the workflow using the Musculoskeletal Atlas Project (MAP). Modeling was image based, it did not use probed points. Segmentation was performed using home grown Matlab script to export point clouds. Statistical shape model was used to predict the bone geometry, which can fill in data that is outside of the scan, as in proximal and distal portions to represent the whole bones.A lofting script capture the cartilage geometry. Thickness parameter for cartilage was incorporated. Overall, the meshing was and an automated process, it took about 20 minutes. Getting ligament origin and insertions correct was where they struggled. Statistical shape model provided attachment sites but these needed to be tweaked to represent the locations in the scan. The team came to the conclusion that the template representation of insertions in the statistical shape model may not be as good and fitting with thin plate spline did not predict the insertion locations as well. Models were run in FEBio and included passive flexion and additional simulations of anterior-posterior and internal-external rotation laxity at 0 and 90. Coordinate systems were based on Grood and Suntay. They struggled to get models converge when using original specifications, ligaments as line elements and lack of meniscus. They decided to move on to continuum representation, leveraging Open Knee(s) - Generation 1 model. The loading and boundary conditions were similar to Cleveland Clinic group. Main protocol deviations were modeling menisci as continuum, modeling tibiofemoral joint ligaments as continuum and update of boundary condition, e.g. tibia fixed. Their overall goal was to reduce burden and automate the workflow as much as possible.

   • Shady described Hospital for Special Surgery workflow. They are not FE modelers, their representations are multibody dynamics models. The team tried to stick to specifications. The have a discretized meniscus model adopted from Trent Guess' approach. They did simulations for contact assessment and prescription of ligament slack lengths, i.e. joint was dislocated and snapped back to adjust for meniscus/cartilage initial penetration. MIMICS was used for segmentation. They needed to convert NIfTI to DICOM. Scans ended up in different coordinate systems after conversion and they needed to perform registration - Ahmet described why he used NIfTI for image data: it is good for analysis, more standardized, doesn’t change based on which scanner is used. When generating geometries, a smoothing constraint 3% from original volume was used. This threshold was experiential. Geometries were in Parasolid format. CT scans were used for DU. They spent most of the time on segmentation. Two medical students followed protocols described in the specifications to segment. Shady did quality assurance, some resegmentation, and still stuck to 3% threshold for smoothing the volume. They thought that it would be a good opportunity to add patellofemoral joint. However they struggled and decided a protocol deviation and add this joint during calibration. They did not use a Grood and Suntay system. Only MCL was wrapped around the bones using via points. In response to Thor's concern (stationary via points may be problematic in deep flexion), Carl described that these points can glide on the bone. They pushed slack length work to calibration. Femur was controlled by 1 dof (for flexion) and tibia by 5 dof. With oks003 simulation took 20 minutes, with DU02 it took 9 hours; surface smoothing and contact were suspects for this convergence discrepancy. Dynamic simulations were performed with ADAMS with a duration of 90 seconds - 1 degree flexion per second.

   • There were a few general discussions. Carl indicated the importance of the origins coordinate systems for description of translations. Reference configuration of the model was kept as in imaging or femur was moved based on residual angles of the coordinate systems depending on the preferences of the modelers, e.g. DU and CSU chose the latter.
   • The participants agreed that there is a lot to do for comparisons between groups: gross comparisons, segmentations, raw and smoothed meshes, coordinate systems, representation type and approach for ligament structures (continuum vs non-linear spring, attachment sites, etc.) and sensitivity on model predictions across groups as a result of simplifications.
   • Development of a high level article for the Model Development phase was found to be necessary. This manuscript can evaluate commonalities and differences in approach, e.g. the ways teams modeled ligaments and meniscus, summarize common challenges and how they were addressed, e.g. overclosure of contact surfaces, explore the impact of different coordinate systems on kinematics, e.g. during passive flexion, and strategies to ensure modeling does not inhibit clinical translation.

2. Model Development Phase Discussions.

   • The teams distributed the workload for publication of a series of articles to decipher "art" of modeling and simulation in knee biomechanics based on the experiences within this project. Ahmet led the publication on overall strategy and he was wondering which groups would like to take the lead to summarize other stages of the project in publication form. Model Development manuscript will be led by ABI. Both DU and CSU teams have an interest to work on Model Calibration manuscript; the decision will be made as work progresses. HSS will work on Model Benchmarking. CSU and DU are interested in Model Reuse.
   • Thor will work on an outline for Model Development article. Potential figures may include each group’s workflow, differences in raw segmentations (mean +/- SD), differences in final mesh (mean +/- SD), registered femur and tibia illustrating the origins and insertions, etc. A potential table may summarize included/excluded structures by each group. A key feature of the study is that five different groups are doing the same thing using the same data but with their workflow.
   • Pete talked about his interest in segmentation results, i.e. align them, overlap them and get differences. Pete will be happy to get involved and connect with Thor to prepare material for the Model Development article. Additional discussions emerged in terms of comparing both segmentations and geometries; start with model ready geometry and go backwards. Evaluation of bones can be a good starting point and later expanded to other tissues. Ligament insertion sites can be compared as well. Yet, the teams decided to leave this after calibration as some participants anticipate to adjust these during the calibration process. Overall geometries may be different but the ultimate question is, does it matter? Activities (simplified simulations) to understand the influence of geometry on mechanics may be pursued, e.g. force-displacement characteristics of ligaments. Carl noted that cartilage thickness and stress is intimately related and demonstration of cartilage coverage and thickness map will be insightful. All these may be highlighted in an overview paper on Model Development, with some examples. Mentioning coordinate systems and kinematics description is a possibility but this may be less interesting as sensitivity analysis of kinematics is well established.
   • Ahmet asked a few challenging questions to the participants: What do you least trust in your model or in each others? What data did you wish you have but was not there? These initiated some discussions. Nynke noted her concerns with cartilage and meniscus contact and ligament attachments and cartilage geometry at the bone interface. Geometry and insertion sites of ligaments were noted as bottlenecks for streamlining. Ahmet wondered if Open Knee(s) data was easier in this regard due to the inclusion of MR images at three orthogonal planes. Some teams used more than one scan, particularly with Open Knee(s) data. HSS team mentioned their other work, indicating that they segment two different scans where they perform quality assurance by segmenting a common tissue. They noted that for DU02 they obtained bone from CT, cartilage and ligaments from MRI. For oks003 bone was obtained from on MR scan, cartilage from another and ligaments from all. Don mentioned volume differences in bones segmented from CT and MRI.
   • Participants also iterated on the ideal message to give to the biomechanics community. Thor emphasized the importance of benchmarks, i.e. using models to answer a clinical questions. It will be good to have a hierarchical evaluation of the impact of modeling art on decisions. This will depend on the research question and context can be created based on importance and relevance. Jason used his teaching experience as an example to attribute differences to assumptions made. As these discussions evolve, future utility of modeling for prediction and extrapolation will likely mature in accordance with mechanical and clinical validity.

3. Model Calibration Phase Discussions.

   • There were some discussions in regard to coordinate systems and how to align each team's processed calibration data and model predictions for comparisons. Ahmet clarified this issue by emphasizing that the Model Calibration phase asks for fit data as interpreted and used by groups.
   • The teams agreed that Model Calibration specifications will be delivered by the end of July. If needed, this will be pushed to August. As a result, delivery of Model Calibration phase outcomes will be pushed to April 2020 (9 months total for execution).
   • The teams planned to discuss Model Benchmarking phase in November 2019 while everyone is half way through Model Calibration. We need to be aware of timeline, as we are potentially 6 months beyond original delivery date. A no-cost extension to project period is a possibility.
   • Ahmet noted that groups can use whatever earmarked data available for calibration, but all calibration data sets should be simulated and delivered by the end of the phase (regardless of whether it was used during calibration).
   • The importance of realizing a calibrated model with the potential to be run by others during the benchmarking and reuse phase was emphasized.
   • If necessary, “intermediate” models should also be documented during the calibration process. It is recognized that “initial guess” models may be intuition based in order to provide a better starting point for the group-specific calibration procedures.
   • Ahmet gave an overview of the earmarked data for OKS data sets. The anatomical and “optimized” joint coordinate systems in the OKS data set were discussed. In this data set the knee the coordinate system can be mirrored since the robot assumes a right knee always (not for every output - look at the documentation). Links are included in the wiki site for OKS robotics testing procedure documentation. The group noted the challenges in processing the experimental data for model use. Specifications will be written with each group’s best effort to describe how this data will be processed. Ahmet asked whether his group’s knowledge of their data should be summarized to help the other groups. At first, we will try on our own and re-evaluate if there are issues (after writing the calibration specifications).
   • Pete noted the DU data is being reverted back to a more “raw” form, which will also require each group to process for model use. Kevin took a few minutes to describe the DU data: passive load versus displacement data for AP, IE and VV between 0 and 120 degrees flexion. It’s all described in Grood and Suntay based on probed anatomical landmarks.
   • As Model Calibration phase progresses, there will be discussions on any problems with processing the data to make sure we’re all (at a minimum) basing calibration on properly processed data. This practice will be valuable to provide insight into the “usability” of the data.
   • The teams are reminded to pay attention to “known limitations” of the data sets, which are included in the wiki descriptions of each data set.
   • Ahmet noted that registration will need to be handled using data-specific procedures, which will of course need to be documented.

4. Other Business.

   • Carl noted that the JOR manuscript is still a possibility. The outline is fairly clear. The content will leverage panelist handouts, flash opinions, and survey results. Carl also went over the survey results with the group.
   • Ahmet mentioned the possibility for supplementary grant application, which can be used to support outside groups to reproduce workflows of individual teams.
   • The participants briefly discussed model curation and certification and how to tackle these.
   • Jason did not have a chance to work on the review paper on reproducibility potential. He is still interested in following up on that. HSS didn't submit their reviews for that manuscript. Assimilation of information acquired from the review of individual papers will be difficult. Jason emphasized that the key aspect is what is reproducible or what is not. The data drive focus. The manuscript will have the potential to provide a reporting guidance while the inability to include everything in a publication was acknowledged. In this regard, Shady learned adding appendix for your methods is better than referring to previous publications that result in layered citations. Jason mentioned that Farshid Guilak from Journal of Biomechanics was interested in and he will send a draft to him.
   • Don presented an overview of slides illustrating the differences in the models from the Model Development specification documents of each team. He will create a Google doc and share with the groups where we can each make sure the specifications are complete, given the protocol deviations. Each group should also incorporate protocol deviations into a final Model Development specification document, which accurately reflects the model that was delivered. These documents should be called "Model Development Specifications Delivered" to differentiate them from Model Development specifications a-priori execution and introduction of protocol deviations.
   • Ahmet noted that the Model Development outcomes of all groups are uploaded to the source code repository. Yet, the files are large. He shared a copy of the files with the teams.
   • In Meshlab, Ahmet opened femur cartilage raw geometries from packages of all teams. An immediate issue was that some groups' geometries appeared to be 180 degrees flipped relative to others. The teams suspect that this is due to interpretation of image coordinate system in different image segmentation software. Nonetheless, Ahmet was able to demonstrate the variations between surfaces using Hausdorff distance.