The Craniobot is an open source tool that allows you to do precise microsurgical procedures on rodent models commonly used in neuroscience research. These models, including mice and rats, can be bred up in the lab quickly and their brains are very similar to human’s. For instance, humans and mice have parts of the brain for processing visual information, the visual cortex. They are good analogs.

Whenever we need to do anything to interface with the brain (for instance to insert electrodes or implant windows on the brain) to image or visualize neural activity, it involves a procedure called craniotomy. Craniotomy is a process where you remove the skull on top of the region of the brain, that you want to access. That can either be replaced with, for example, a glass coverslip or it could be a burr hole through which you insert an electrode. If you want to do any invasive neuroscience experiment, that is the first step you do. It’s challenging because the skull of a mouse is really thin in the range of 50 micrometers to a few 100 micrometers depending on where you are.

Traditionally, craniotomy has been done by neuroscientists using tools that are adapted from dentistry. It’s similar to what your dentist uses to chip away plaques around your teeth. With this tool we need to do it very carefully. You want to only remove the bone and not damage the brain underneath. This was a nice process for us to automate. It’s a challenge for us and a lot of other neuroscience labs. We come from an engineering background. If you have a sense for how thick the skull is at a particular location and what the contour of the skull is, you can program a computer numerically controlled mill or some kind of robotically guided tool to precisely chip away the bone. All of this is done without shaking hands. That would be otherwise hard to do for beginner surgeons. There’s a lot of skill involved. That is what motivated us to work on the Craniobot.

It removes the damage to the brain completely. We’re talking about a very delicate procedure. There’s skill and speed involved. The faster you can complete the procedure, the better it is for the animal’s health later on. The skill factor is relevant because everyone is different. Your hands are usually shaking during a surgery. Humans are not very good at performing precise motor sequences or activity repeatedly. If I ask you to follow a trajectory in 3D space multiple times, there’s going to be a lot of jitter. That’s precisely what you need to do for these kinds of procedures. That’s why a robotically controlled cutting tool is very useful. It takes away that variability in human motor activity.

Initially, it was all funded through lab startup funds. This is one of the first projects we took up. If you start a lab at any university, usually you end up with some startup funds. Most of this was funded through those funds. Later, we had support from NIH. NIH is the National Institute of Health here. I guess it’s the equivalent of the European Research Council.

It was a much larger grant for a bunch of related activities, because the project didn’t cost a lot. The whole objective was to build tools that can benefit the neuroscience community. You don’t want to build anything that’s sophisticated and needs a lot of monetary resources. If you look at lab maker, they’re selling it for about $3,500. It’s a lot. Most of the grant money was used to support graduate students through their PhDs for their work. Graduate students in the US cost about $45,000 to $55,000 a year with stipend, tuition and health insurance. We were funded for a certain period of time for a student to make this. This project took about two years from the ground up.

Our objectives from a design standpoint were very simple. In the past, I had worked and helped on a related project, when I was a postdoc. We tried to do the same thing through impedance sensing. This means to recognize when a probe is going through the skull. When it does, the impedance drops. We tested here a force sensing method, which we thought was more reliable. We wanted to build something that was built for machining rather than something that’s integrated into an existing surgical apparatus, in contrast to what previous architectures were exploring. This allowed us to take a CNC mill that’s used by hobby woodworking enthusiasts. Our hobby woodworking enthusiasts use sustainable top CNC. You can buy it from Amazon for about $500. We want to do a force sensing based method. One of the innovations we had to figure out, was to build a contact sensor that could measure contact with the skull surface accurately enough, while applying very very little force. The mouse skull is soft, it’s not a rigid structure. That was the primary innovation. Once we had that, we could leverage all the CNC routines and other things that people typically use in robots to control the machine. The impedance sensing method wasn’t reliable for a number of reasons. We thought this force sensing approach is a more robust and alternative approach and that’s what we wanted to do.

I was the primary supervisor of two students, Dr. Leila Ghanbari, who is now graduated with a PhD, and Matt Rynes, who is still in the group. They worked together. Leila mostly worked on the engineering aspects, while Matt worked on the animal research. That was the bifurcation.

There were a few bottlenecks. One bottleneck was focus on inexpensive mills. We wanted to make this affordable from the ground up. That meant that the mills aren’t very accurate. It’s not a sophisticated machine shop tool with the same accuracy. These mills have drifts. Because people will try to use this for very delicate animal procedures, we needed to be careful. We had to iterate a bit in that space. There was a trade off, you couldn’t just buy the cheapest CNC mill available. We used quite a bit of time dealing with this issue.

We bought off the shelf parts. Some of them didn’t work out because, from a control perspective, they weren’t accurate and repeatable enough. We were trying to do this in animals, that added complexity. Some of these evaluations are much slower than if you’re dealing with inanimate objects. The iteration cycles are quite slow.

Further, anything biological has high variability. Machine tools, that are designed for accurate machining of inanimate solid blocks of metal, are different from machining a soft living tissue.

A lot of the components are off the shelf. The innovation came in from the system integration side. The main integration would be how to do contact sensing of the skull using very small millimeter scale forces. You accurately determine what the skull surface is and make use of that information. The other innovation is determining how to make use of the information: how deep do you need to go into the skull and in what regions? It’s a lot of experimental heuristics.

It was a mixture of both. It has to do with the PhD-student academic advisor dynamic. When we started the project, both students were early in their training. This was more authoritative. As the students became more independent, the decision making was more democratic, like in any mentor-mentee relationship.

You can break it down to three main components. There’s the CNC mill itself, that is off the shelf. We had to develop a specific apparatus that allows to anesthetize these mice or rats and fix them securely within the CNC mill. This is called a stereotactic apparatus. We needed to build one that’s customized to fit into a programmable CNC mill device. The third component is the contact sensor.

I would say the CNC mill itself is an electromechanical component. The stereotax is primarily mechanical. It’s a structure for securing. There’s nothing electrical. The contact sensor is a combination of precision engineered mechanical elements and sensing electronics. It’s electromechanical.

There’s a software that we developed. It’s written in Python and MATLAB and does the overall control of everything.

It’s somewhere between a refined, well put together laboratory prototype and a turnkey. If you look at what LabMaker has done, they have a turnkey product now, that they’re selling on the website. We spent about six to eight months with their engineers. Besides, people can take the designs that we have published and put on GitHub and other repositories. There have been a few labs that have independently got this system to work.

A few groups did it. York Winter’s group is one example. They did a few design improvements and iterations to make the contact sensor more robust. They used different spindles. The tool is rotating more precisely and with less noise. There’re a few iterative design improvements that York Winter’s group did. They were looking at it from a standpoint of making it more turnkey. They have created nice hardware interfaces. A second group is at Rice University. They improved this control software and they wanted to adapt it for a more specialized procedure. These are two examples which show the efforts. There are other groups who have taken it and modified it for various other applications.

Some modified it independently and I got to know about it from conferences. In some cases, they actively got in touch with us.

The primary objective was the publication, but also building a tool that would be very useful in our lab. I didn’t want new grad students coming in and spending months learning how to train and to do these procedures. I knew if we solved this internally, it could be a widely applicable tool. A lot of the open source push happened because of the response we got from the community when we started presenting this work. That was when some of the conversations with the York Winter happened. They were interested in taking open source tools and making turnkey products, that they sell out of their LabMaker outfit. This way the collaboration came about.

We realized that the initial prototype needed a bit of refinement, before we can start supporting other groups. Taking a laboratory prototype and providing those designs by themselves can be challenging, because there’s a lot of troubleshooting. We spent a lot of time working with the early beta testers. We took the time internally to build a refined version of the Craniobot, which we published as a follow up at Nature Protocols. It was a multi-step process.

There was no sudden finishing moment. It was a slow iterative refinement, until it worked really well. Afterwards we thought about experiments to show that it works well enough.

Everything is accessible. We published a detailed, step by step protocol. It contains, for example, the CNC machine and what you need to adapt. We have a detailed protocol on how to put the hardware together. The protocol covers how to do initial testing in the benchtop, before you do surgical procedures on rodents safely and in a sterile fashion. All updated designs are available on our GitHub repository, whether it’s software or CAD files.

Github is the primary repository that we use. We published the hardware in the Nature Protocols paper.

I think GitHub is mostly easy to use. We use Github internally to share files and repositories. Nature Protocols is the best place to publish procedural documentation. It’s a nice avenue and widely read by biologists as well as engineers.

No, we perceived no barriers on these platforms.

Sure. It’s always a learning process. What’s the best way to disseminate the tools? What’s the best version? Do you continue refining? Is the software ready? Every group and end user is different. As an academic research lab, there’s a huge bandwidth on tool improvement and refinement, for which we don’t get academic credit. It’s always a balance. How much do we need to take this further? If we take it further, what are the subsequent benefits to us and to the community?

We didn’t specifically think about any other. We tried publishing the preprint of the protocol on an online repository, like bioRxiv. BioRxiv has a policy of not publishing protocols, which is an interesting barrier. BioRxiv is about open science. They rejected us, which is the only rejection I’ve ever had from BioRxiv now, because their policy is to only publish new results or other kinds of prospectives. I don’t think they publish review articles.

We have published everything, after the Nature Protocols paper, we worked with York Winter’s group. This is as much as we can do.

We’re using it regularly within our group. We work in an academic lab and we’ve moved on to other projects. We help other open source tools that we’ve developed.

Our lab and at least a few labs will benefit from the project. It opened up a whole bunch of other approaches. It opened the idea that you don’t need to have sophisticated tools to do this kinds of automation. It takes time and work, but it’s possible. That would be one thing I can point out. The second thing is the benefit to the community. There are quite a few groups that are building on it and automating other aspects of implantation. One direction, that we’ve taken, is to discover more precise, without contact and non invasive sensing of the skull thickness. I guess, it would be hard to make that open source, because it involves much more expensive optical coherence tomography and imaging. That’s something that we’re actively pursuing. It’s an extension from a different approach. I don’t know whether it will result in an open source tool, like we had envisioned when we originally started off.

It’s about the costs. We’re trying to figure out if we can non-invasively and as a non contact fashion determine the skull thickness, because right now we’re going in blind. That involves using OCT imaging, the imaging system by itself is pretty expensive. We’re talking about €20,000. We could make it open source. I don’t see any reason why we wouldn’t. I think there’s a difference between building an open source tool that costs €5,000 versus €30,000. The benefit of open sourcing is probably lost in those price ranges. We go for something that’d be open source but accessible only to a few select labs.

For example, there’s an open source, two photon miniature microscope that’s been developed by the Moser group in Norway. And there’s a single photon miniature microscope that’s been developed by the UCLA group and a whole bunch of others built it. The two photon miniaturized microscope is extremely useful in terms to the value proposition, resolution or other capabilities that it brings. However, it requires a two photon laser, which costs €150,000. It can be open source, but it is going to benefit a significantly smaller number of labs, compared to an open source microscope that doesn’t involve such a laser.

There was a clear need within my group that I needed to solve. Like I said, solving this for us meant solving it for a whole bunch of other people. We were in the right place and time to do it, because I straddled both disciplines as a neuroscientist and predominantly running an engineering group. We had the necessary skill sets to tackle this problem.

Quite a few people assisted us who played supporting roles. Daniel Schulman did a lot of coding. Sarah did refinements on the surgical side. Greg Johnson worked on the mechanical hardware and structural engineering. Michael Laroque contributed to the mechanical design.

It’s been 8 people in total.

They contributed later on. Especially, Labmaker helps us. They have experience in taking lab prototypes to market ready products. Some doctoral researchers supported us from Rice University.

No, we didn’t need them.

Jay was a master student. Matt was a master student and now he’s a PhD student. He was doing his master’s thesis. Daniel Schulman was an undergraduate student.

They just congregated. We didn’t recruit them for this particular project. We have a good engineering department who have a lot of the skill sets, that are needed for this project. I had expertise about neural engineering and animal work, which is what I train them to do.

The engineers had specific tasks to do. Someone had to do the lab recording. Another one needed to do the animal research. Often, these tasks were staggered across time. Different people came in at different times and did the needed improvements, like refining a particular structure or mechanical part. At any point, there was probably one student interfacing directly with me.

We’ve thought about it. It would be useful in humans. We’ve explored that a little bit, but we have nothing active at this point.

Yes, definitely. In terminal experiments, the mice are euthanized at the end of the procedure. If it’s a survival surgery, you need to close up the animal and there are some procedures to do that safely and in a sterile fashion.

For Leila, who was a PhD student, this is a significant part of her PhD thesis. This is generally true for most trainees in my group. They primarily come from engineering backgrounds and get to work in a very interdisciplinary environment, which is tackling instrumentation problems for neuroscience. That translates to a lot of work in the medical devices industry. Therefore, you need people with the engineering skill sets and quantitative skill sets to be able to speak the language of clinicians and biologists and come up with solutions in that domain. Leila directly benefited from this by becoming a research engineer in a big medical devices company. Matt, who was co-first author, won the University of Minnesota distinguished Master of Science thesis award for his work. That benefited his career. He has won a number of other fellowships since then. But this was the launchpad. Michael Laroque went from engineering to medical school and he tries to be a doctor. Daniel Schulman did a lot of the mechatronic aspects of the project as well as the coding. He’s now a grad student in Michigan, where he’s doing very similar things, but at a much finer and more precise application.

This tool is the very first paper that came out of my group. I benefited hugely, because I learned how to direct a research project and publish a paper, which is a huge learning curve for someone who’s starting their own research group. Everything I did previously was under the supervision of somebody else. To lead something was the biggest benefit.