Deep learning algorithms and AI are rapidly being adopted in medical imaging applications due to the increased demand on radiologists and medical experts, resulting in undue stress, interpretive errors and a general need for a larger workforce. In fact, there’s currently a large error rate in the interpretation of PET/CT scans, resulting in ~40 million errors per year. This rate of error has remained virtually unchanged for years.
It’s widely known that PET/CT scans play a critical role in clinical applications in oncology, including aiding in cancer diagnosis, staging, assessment of treatment response and surveillance. The hope, with deep learning and artificial intelligence (AI), is to improve diagnosis and reduce the demand on medical experts.
Last year we were fortunate enough to collaborate with leading researchers at Stanford University’s Department of Radiology to conduct research* to just that: apply AI to brain PET/CT scans.
Using machine learning with our NeoPulse® platform, the goal of the research project was to see if we could train and evaluate the accuracy of a computer algorithm to detect abnormalities in PET/CT scans. Put simply, we wanted to determine if AI could conclude if a scan was considered normal or abnormal at a similar or better rate than a radiologist.
If we could achieve that, we knew we’d be able to improve the efficiency of the overall PET/CT process where a radiologist goes from taking an image, to reading, to diagnosis and then treatment. The problem with this current state is that the bottleneck in this process is the radiologist. If we can automate the process, we can reduce their workload and improve their efficiency while getting faster care to those who need it most.
Ultimately, we wanted to able to show we could build a model that can be used in conjunction with radiologists to give them an efficient way to flag certain scans with an algorithm that basically says ‘hey you should pay more attention to this image and not worry about these other ones.’
So, to kick off the study, we build two-dimensional convolutional neural net models for various window settings. Let me briefly break down what I mean here.
If you’ve ever had a CT or PET scan, you might know that you get a number of images that all display different angles or views. Even if you haven’t had this experience, you can imagine one angle may not show anything, and another may. You then have radiologists who are going to read those images manually and make a determination as to whether there is cancer or some other abnormality in the image. That’s a hard problem for even a specialist to handle.
Ok, back to the study. A total of 289 patient anonymized scans (totaling 68,260 individual images) were included in the research. The brain scans were then separated based on binary outcome, as either normal or abnormal based on their clinical reports. We randomly selected 100 normal and 100 abnormal for training and validation, then used the remaining 89 for a test set.
And, just like a radiologist would do, we trained the models for a variety of different ways to view the scans. We then calculated the probability that the scan was normal or abnormal and took an average over all the models. As you can imagine, bringing all the images together significantly increased our accuracy. We were thrilled to achieve 82% accuracy with the NeoPulse® platform and posited that similar models could eventually provide decision support in a clinical setting.
This is a big breakthrough to see how the power of AI can assist medical professionals to help not only reduce the burden on them but allow for them to exponentially scale their expertise and provide more rapid and efficient medical care to more people.
However, if you’re a scientist or engineer reading this, I’m sure you’re saying to yourself that the dataset is far too small to be confident in rolling this out to clinics around the world just yet. And you’re right. I’ll be the first to admit that this is just an initial step along the way – albeit a very promising one.
When this approach is validated with a larger core of patients, we are well on our way to clinical support and reducing the workload on radiologists significantly
Here’s what that might look like: An app in the radiology room where the scan is performed that takes the study and runs it through an AI model that then gives the radiologist an indicator of whether they need to spend extra time looking at certain images.
We believe this study is groundbreaking in showing, not only the power of NeoPulse®, but how AI can rapidly improve healthcare for everyone involved.
– Jason
*The final publication is available here.
Nobashi, Tomomi & Zacharias, Claudia & Ellis, Jason & Ferri, Valentina & Koran, Mary & Franc, Benjamin & Iagaru, Andrei & Davidzon, Guido. (2019). Performance Comparison of Individual and Ensemble CNN Models for the Classification of Brain 18F-FDG-PET Scans. Journal of Digital Imaging. 33. 10.1007/s10278-019-00289-x.