Headquartered in San Diego, California, 3DT (Drugs and Devices for Diagnostics and Therapeutics) Holdings is an incubator for technology-based startup companies in the fields of cardiovascular and metabolic diseases. By providing expertise in pathology, bench studies, computational modeling and clinical research, 3DT Holdings contributes to the discovery of new and useful devices and therapeutics that enhance patient outcomes.
To help improve patient outcomes for a specific form of heart-valve disorder, 3DT Holdings is building a decision-support solution that will deliver real-time guidance to surgeons in the operating room. As a first step, 3DT Holdings created a model for this surgery and set out to train (optimize) the machine-learning model — iterative simulations — based on as much real-world data as possible. By partnering with high-performance computing (HPC) experts at UberCloud, 3DT Holdings was able to use a containerized environment on Google Cloud Platform spanning more than 3,000 Kubernetes clusters — all supervised using Rancher Prime. The SUSE solution automates repetitive processes, facilitates management and ensures high simulation success rates — all of which minimizes operating expenses. This empowered 3DT Holdings to train its new model within budget and move to the next phase in product development.
Helping surgeons through machine learning
Cardiovascular disease is placing a growing burden on health and care systems worldwide. 3DT Holdings develops technologies and techniques that can help clinicians to intervene more effectively.
One condition that 3DT Holdings aims to tackle is mitral regurgitation, which affects the mitral valve — a small flap in the heart that prevents blood from flowing the wrong way. Without monitoring and treatment, mitral regurgitation can progress from mild symptoms, such as chest pain, to serious complications, such as heart failure.
For patients who cannot tolerate an open-heart surgery, physicians can attach a small device — a MitraClip — to the mitral valve, which helps it to close properly. Surgeons attach the MitraClip via a catheter inserted into a vein, which is then guided to the heart using an attached camera.
Since the size and shape of the heart and the mitral valve can vary significantly, some patients may require multiple surgeries to place the device in the optimal position.
Based on literature, nearly 29% of patients who go through the procedure have complex valve anatomy, and up to 40% of those cases require multiple interventions to reposition the clip. If surgeons can be empowered to place MitraClips in the optimal location on the first attempt, it will make a powerful difference to outcomes and experiences.
3DT Holdings recognized an opportunity to harness echocardiography data from previous surgeries to give surgeons real-time guidance on the optimal placement of MitraClips. Creating this personalized medicine solution required the organization to build a machine-learning model and train it on real-world echocardiograms taken before and after surgery.
“Our work with UberCloud and SUSE is playing an important role in developing an innovative solution to heart-valve disorder, one of the healthcare industry’s most pressing challenges.”
More simulations, greater accuracy
While 3DT Holdings initially developed its model on a workstation, the high compute resources required to run the actual training simulations were too much for conventional desktops.
Using funding from the federal Small Business Innovation Research (SBIR) program, 3DT Holdings engaged HPC specialist, UberCloud, to help move its simulations to the cloud.
Yaghoub Dabiri, scientist at 3DT Holdings, says, “In effect, we challenged UberCloud to minimize our cost-per-simulation so we could maximize the training dataset for our machine-learning model.”
UberCloud proposed a solution based on Google Cloud Platform, using Kubernetes clusters running on compute optimized (C2) preemptible virtual machine (VM) instances. The simulation software is Abaqus Unified FEA from Dassault Systèmes. UberCloud also deployed Rancher Prime — an open source platform that allows organizations to deploy, manage and protect Kubernetes clusters at speed and enterprise scale.
Daniel Gruber, director of architecture at UberCloud, comments: “The crucial element in the solution is Rancher Prime. Preemptible VMs are a great fit for 3DT Holdings, reducing cloud HPC costs by around 80%. However, these instances do not offer guaranteed uptime and can only stay running for a maximum of 24 hours. To eliminate the risk of simulations being shut off prematurely, we had to run each simulation in its own Google Kubernetes Engine cluster and automatically restart it at the end of the process.”
He adds: “Supervising up to 3,000 Kubernetes clusters manually using vanilla tools would simply be impossible. Rancher Prime provided the automation, monitoring and management that made the solution feasible.”
Deployed on an UberCloud environment in the Azure cloud, Rancher Prime is now fully integrated with the Kubernetes clusters on Google Cloud Platform, empowering the organization to automate key tasks, such as applying Kubernetes updates as well as general monitoring.
For example, Rancher Prime allows UberCloud engineers to get a real-time, fine-grained view of all simulations in progress. Equipped with this insight, the organization can quickly identify and resolve issues, such as underutilized compute cores or suspended clusters.
“Thanks to the combination of cost-effective C2 instances and a consistent application stack that we can manage using a single command in Rancher Prime, we enabled 3DT Holdings to maximize its cloud resource and license usage,” explains Gruber. “With Rancher Prime, UberCloud helped 3DT Holdings to hit its target of training its new machine-learning model using more than 3,000 simulations.”
Wolfgang Gentzsch, president of UberCloud, adds: “Our collaboration with SUSE empowers UberCloud to provide tailored next-generation HPC solutions that offer a perfect fit within our customers’ enterprise environments.”
Taking the next step
Through its partnership with UberCloud, 3DT Holdings has achieved the first important step in the development of a brand-new solution for personalized treatment. The company has now successfully created a proof-of-concept machine-learning model capable of predicting the outcomes of possible MitraClip placements — even for patients with complex valve anatomy. The solution won the prestigious 2021 HPCwire Award for Best Use of HPC in the Cloud.
The ultimate aim is to develop a real-time decision-support tool that will instantly show surgeons in the operating room how effective the placement of the MitraClip is predicted to be for any given location on the valve.
As a next step, 3DT Holdings plans to build a production-ready version of its solution, which it will ultimately submit to the U.S. Food and Drug Administration for approval.
Dabiri comments: “Our vision for the solution is to allow surgeons to take ‘snapshots’ of echocardiograph scans during surgery and send them instantly to the machine-learning algorithm for analysis. In a matter of seconds, the solution will show predicted strains and stresses as red dots overlayed on the scan of the valve. The surgeon can then make slight adjustments to the position of the clip and repeat the analysis until red dots and leakage flow are minimized, indicating that they have found the optimal position for the implant.”
Looking ahead, 3DT Holdings sees Kubernetes clusters as the ideal platform for its production-ready solution — and the company plans to leverage its partnership with UberCloud and SUSE to help bring the concept to market.
“Our work with UberCloud and SUSE is playing an important role in developing an innovative solution to heart-valve disorder, one of the healthcare industry’s most pressing challenges,” concludes Dabiri. “We’re looking forward to launching a new generation of personalized medicine solutions with the potential to improve quality-of-life for millions of people around the world.”