I spent a couple months after completing my Master’s working on creating performance tests/metrics for a novel unicompartmental knee arthroplasty robot. Now, for obvious IP reasons I’m not comfortable showing photos of the robot I was working with or sharing many details about it since it’s currently in development with a group at UBC. However, I can share with you attributes of the protocol that I developed.
As is the case with many startups, you need performance data to lure investors but you also need cash to develop you device to the point that you have good enough data to lure investment. A tough cycle to break, particularly when you’re lacking the initial capital to get things off the ground. During my Masters I became very good at simulating and performing experiments without a lot of investment, mostly because it was a great opportunity to build an entire experiment from the ground up, starting with an idea and ending with a thesis (and pending publication). I took these skills and applied them to this project to great effect.
The golden rule of computer-assisted orthopaedic surgery is that successful interventions are ones where the implantation error, the difference between the final implant position and preoperative plan, is both sub-mm in translation errors and sub-degree in rotation errors. From the beginning of this project, I knew I had to be able to measure errors on the order of a 1/10th of a mm in order to achieve my goal of measuring the accuracy of the implant position. I would discover that without a dedicated machinist, that’s really hard to do!
The first two deliverables of the project were a clamp to hold the robot relative to the simulated anatomy, and the simulated anatomy itself.
Using constraint theory, the clamp is capable of positioning the robot with sub-mm deviations provided the nuts are tightened to a similar level each time. The oak platform has a v-groove routered into it to allow the mounting rod of the robot to be clamped in place by the flat UHMWPE bar, while self-centering relative to the simulated anatomy. I placed it at a 45 degree angle to simulate the typical limb position during surgery as a means of assessing the robot’s workspace. The block in the foreground simulates a femoral condyle, with dimensions based on literature values of existing UKA implant dimensions, and anthropometric data on distal femoral condyle dimensions and radii of curvature. Held in place against the front aluminium bar using a thumbscrew to enable rapidly changing samples during experimentation, it’s simplistic shape was chosen to provide the capacity for rapid iterations on cut-surface shapes and was manufactured by a local machinist’s CNC table router with an accuracy of +/- 0.05mm. The arch on the lower portion of the block allowed for kinematic registration to the main axis of the robot for repeatable positioning.
Once the physical components had been created, the group needed a method of assessing implant accuracy relative to a preoperative plan. I had spent quite a bit of time creating and validating a laser scanner -> Rapidform workflow during my Masters and saw an opportunity to apply these same tools here. Since the main axis of the robot was used to align the anatomy, it made sense to use the same kinematic registration to the laser scanner coordinate system. A kinematic registration rig was designed in Solidworks and CNC’d using a Roland MDX-540 (New toy at the research centre. Had to learn how to use it!). This jig used mounting pegs to align itself on the rotating stage in front of the laser scanner, ensuring that the samples, which were kinematically registered to the jig, were also registered to the stage.
Samples were scanned using a VIVID 9i laser scanner. Previous analysis has shown that we were able to detect differences on the order of 0.05mm. Two outcome measures were reported for the group.
Surface Roughness Error Procedure
Scan data was registered to the CAD model using the alignment pegs on the kinematic registration jig, allowing the scan information from the samples themselves to be excluded.
Once registered, the surface of the cut model was extracted and a mesh deviation between the cut surface and the CAD surface was performed, calculating the average gap distance (deviation) and RMS error. It was observed that an asymmetry existed between the anterior and posterior portions of the surface, so the error metrics were calculated again and split into halves to highlight this. This was achieved by distributing comparison points across the surface, spaced by 1 mm and 1̊, for approximately 1200 reference points per side.
Finally, an ICP algorithm was used to minimize the deviation between the cut surface and the CAD model to analyze the impact of surface roughness on implant alignment. In order to ensure that the algorithm actually minimized the error, a mesh deviation was performed and the RMS error was recorded. The transform from the ICP algorithm, called the “best-fit” transform, was saved and used for implant alignment analysis later.
Implant Alignment Error Procedure
Scan data was registered to the CAD model using the landmarks on the kinematic alignment fixture to avoid losing relative alignment of the implant to the ground truth, as discussed above.
The transform between the scanned sample and the CAD implant was obtained by placing a local coordinate system on the corner of the CAD implant, and manually aligning the scanned data. The manual alignment is performed by combining rotations and translations in the respective axes, by hand, until the implant appears to be aligned. This manual alignment process was briefly assessed for single-observer repeatability/accuracy, with rotation and translation resolution being approximately 0.3̊ and 0.3mm.
The recorded outcome measures for this procedure were RXYZ and TXYZ, where X (red) follows the length of the implant, Y(green) follows the height, and Z(blue) follows the thickness of the implant. For clarity, this error metric is the transform between the CAD reference implant and the scanned implant, subject to both surface roughness/distortion errors, and bias (rigid body errors). The previously mentioned “best-fit” transform was then applied to the scan data, and the RXYZ and TXYZ analysis was repeated.
So after all of this effort, running dozens of samples for statistical power, how’d I do? Splendidly. Errors were sub-mm and sub-degree, often bordering on the limits of the testing sensitivity. There are better ways to perform the analysis I’ve detailed above but CT scans and cadaveric models are expensive. When you need quality proof of concept work for not a lot of money, I think this is going to be tough to beat.