There’s an ambitious initiative underway to bring 3D ultrasound scanning, which uses echoes from different angles to create a complete 3D image of a body part, to medical clinics everywhere. Called CUDL (Collaborative for Ultrasound Deep Learning), experts across the world are working to create a cloud-based service that will allow doctors to upload ultrasound scans for interpretation. To support it, the CMA’a Joule recently awarded CUDL funding to develop its technology. In this Q&A with Dr. Jacob Jaremko (pictured), who leads the Canadian wing of CUDL out of the University of Alberta in Edmonton, we explore how 3D ultrasound can benefit doctors and patients globally.
Q: Why do ultrasound images need interpretation and assembly into 3D images?
A: Doctors typically get a noisy blurry picture when they scan a body part like the heart. 3D ultrasound processing is like processing an out-of-focus image taken at night of the heart and sharpening it with computer enhancement to make the image look exactly like a heart. This tackles a growing issue: Doctors are taking ultrasound images but often don’t know what to do with them. That is the main problem with ultrasound right now. Everyone can take a picture with the ultrasound but nobody knows how to read the images with confidence to make a diagnosis.
Q: What was the impetus for joining the CUDL collaboration in 2015?
A: The CUDL concept in general is to try and get the most out of ultrasound and to make it a more useful imaging modality for physicians everywhere. We think one way to do that is by a deep learning approach where you upload the ultrasound scan that you’ve obtained and get it read by a computer program that has analyzed thousands of other images of the same body part or the same area and is learning from that vast experience to recognize what is normal, abnormal or important.
In Canada, our best imaging modalities today are MRI and CT scans but those are expensive machines that aren’t readily available to many patients, especially in remote areas. But ultrasound is becoming incredibly compact and portable now. It offers the possibility of being a real point-of-care imaging test that you could think of as the 21st century stethoscope. However, some of the impediments to that are pictures are blurry and grainy and difficult to interpret, and the quality of the scans depends a lot on how experienced the user is. Some of these problems can be dealt with by using 3D ultrasound.
Q: Would doctors need to get high-end ultrasound scanners or other equipment to access the 3D ultrasound cloud?
A: No, you wouldn’t need a fancy 3D scanner in your clinic, you would just use regular ultrasound probes, which are rapidly becoming much cheaper. Scanners cost a few thousand dollars now but they’ll probably come down to a few hundred in the near future. We’re working to allow doctors to, for example, log into their Joule account at the Canadian Medical Association and upload their patient scans. The service would provide answers based on the computer analysis to sort out which scans look normal and which ones need further examination.
Basically, it’s a computer decision support system. The physician still makes the real decision but this is with the experience gained by a computer system that has learned how to detect abnormalities by looking at thousands of scans.
Q: How accurate is 3D ultrasound compared with MRIs?
A: My role in this academic collaboration is in fact to help with the validation. We published a paper comparing the ultrasound of models that we generated of 3D hip surface shapes to MRI. We were within one millimeter every time. We need to do more of those kinds of studies with the best scientific methods so we can validate that we’re correct. We need to follow lots of patients to see what their long-term outcomes are to make sure that our predictions of whether they had, say, a normal or dysplastic hip, or whether they had a tumor or not are correct.
Q: What will be the benefits of CUDL once this cloud-based service is a reality?
A: It will provide family physicians and the people in smaller communities the ability to get easier access to tertiary care and advanced decision making. The practical implication of that is that some things that used to require tertiary referral might not need to. And because it’s inexpensive, we could screen everyone for certain conditions, for example hip dysplasia, and it won’t require a whole network of multiple tertiary ultrasound studies and referrals back and forth. It can actually save money for the healthcare system by having an immediate and accurate diagnosis in certain settings.
For more information, visit http://www.cudl.ai/