Showcase: Riverain Technologies
Part of the Industry Showcase Series : demonstrating AI applications in the imaging world.
I had a very enjoyable (and somewhat lengthy) chat with Jason Knapp and Steve Wornell of Riverain Technologies a few weeks ago. Steve is the CEO and Jason their Chief Science Officer. Riverain is a midwestern tech firm near Dayton, Ohio and both Jason and Steve have deep roots in the industry. How deep? We started out talking about decision trees (Markov chains). They are data nerds, in the good sense. They were around for the first round of CAD-e tools which were incorporated into mammography systems (2nd look and GE’s Digital CAD product) which I believe incorporated simple neural networks & Support Vector Machines.
Their products include a suite of FDA-approved tools for both image processing and anomaly detection. They have four FDA-approved X-Ray products: the ClearRead X-ray Bone Suppress algorithm, Confirm algorithm (contrast/edge enhancement), Compare algorithm (subtraction), and Detect (nodule/mass CAD). They have two CT products, ClearRead CT Vessel Suppress, and CT Detect (nodule/mass CAD). For imaging professionals, note that the CT Vessel Suppress does not appear to be the same as a Minimum-Intensity (MinIp) projection. CT Detect is being marketed as a “concurrent read” productivity tool for radiologists, and the first lung CAD that identifies semi-solid or ground glass nodules. A journal article in the AJR using CT detect can be found here.
Jason and I had an extended talk, principally about their design methodology and how they solve machine learning problems.
We began talking about the book The Predictors by Thomas Bass – about a quantitative hedge fund firm started by physicists & later acquired by UBS that used principles like chaos theory/fractals to (reportedly) beat the market. Jason and Steve are similar, in that they use domain-specific knowledge from physics and engineering to perform “data conditioning” as part of their product design.
Data Conditioning as Riverain describes it, is “bringing to bear prior knowledge” of real-world physics to engineer their solutions. Basically, they are acknowledging there are constraints based in the laws of physics and variation in parameters of the devices that we use for medical image acquisition. Part of their “special sauce” is to bake this knowledge into their system (as data conditioning) to deal with study acquisition variation before it goes to their classifier. Variation in imaging studies caused by maker, model, and modality is ever-present, and an important point likely to be missed by folks without subject matter experience. It is one of the reasons why you don’t see texture analysis of medical images used “in the field” – data from one machine may not be equivalent to another, particularly across vendors, but even between models within a vendor family. This is a real issue that folks working in the MRI segmentation and classification areas are intimately aware of. At the 2017 C-MIMI, the MGH team’s breast classifier initially wasn’t performing well when taken from training to real-life use because of the difference in dynamic range upon acquisition from another vendor’s machine had not been accounted for in the training set.
Jason and Steve feel that Data Conditioning is a large part of the solution and involves data diversity. A side benefit of data diversity is increasing the robustness of the algorithm. Jason explained his approach as “inverse modeling” that they used for the CT CAD product, which is built off the vessel detection algorithm. It uses data synthesis. Typically, image data augmentation in deep learning relies on collecting lots of data and then performing multiple affine transformations, random crops, and the like to increase the effective number of samples, essentially regularizing an ill-posed problem or function. For a mathematically well-posed problem (after Hademard) getting into an inductive proof, one can synthesize additional data that is statistically relevant. Riverain’s approach is to do just that and focus on meaningful, high quality training data.
Jason feels (and I would agree) that these simulated lesions improve the CAD’s performance, and they pay great attention to the representative nature of the lesions they are simulating. He is concerned with the statistical variation of the augmented dataset, and the strength of the samples. This is because of the performance boost that comes with data augmentation in any deep learning model. I would concur: CIFAR-100 error for a 110-layer ResNet without data augmentation is 44.74, but with data augmentation is 27.22. If you don’t believe the impact or utility of data augmentation, you need only look at the NVIDIA results generating fake celebrity headshots (from celebrities that don’t exist) from a Generative Adversarial Network (GAN).
Finally, we discussed some of the practical aspects of running a CAD detector on diverse PACS systems, enterprise systems, and different IDN’s (Integrated Delivery Networks – the healthcare administrator’s name for a hospital). He brought up their philosophy of using model compression to make it run fast. In a world of cloud-based solutions and productivity, I think it is important to recognize how fast a practicing radiologist works and for successful implementation of a product this is important to consider. I haven’t come across this before – but as deep learning moves more from academic pursuit to actual implementation, the inference & edge computing part of development becomes more important.
I very much enjoyed our chat, and wish Riverain much success. I gather they would love for some talented folks to come join them in Ohio.
Remember: If you or your firm/startup would like to be featured in Showcase, you need only ask. Shoot me an email at c.o.n.t.a.c.t@n2value.com (remove the periods) or find me on twitter: @drsxr .
Full Disclosure: I do not have any financial relationship with Riverain at the time of this writing, nor have I been promised any remuneration (or anything else) for this post.
