Today we are living through one of those heady situations in which scientific, technical, and commercial frontiers all simultaneously advance in a grand interrelated dance. Advances in computer technology in the last decade opened up the potential for big gains in applications of neural networks aimed at recognizing and diagnosing visual images. Many startups and established firms are making decisions about how to develop and deploy such software, and what products to develop next.
While the science continues to expand into some headline popping territory, and the technologist find new prototypes for mind-altering forecasting, what has happened within markets? What has found its way into practical applications with commercial appeal?
We can start with a simple observation: If real people are spending real money on an application of machine learning, then it must have created new economic value. However, it can be difficult to provide an overview of a broad moving target.
Today’s column goes inside the actions of one firm’s experience as a way to provide insight into the broad trends. It will focus on the factors shaping the creation of value from applying machine learning in medical imaging. The focus will be on a firm called Zebra Medical Vision, an Israeli startup in the business of using machine learning to improve x-rays. Its experience illustrates many of the same technology trends and strategic issues faced by other machine learning startups, particularly those in healthcare markets. (See references for the longer write-up).
You probably have not heard of Zebra, even though it is doing well by the norms of its field. It is a startup competing with other startups and major technology companies. Zebra’s product is a series of software tools that use machine learning algorithms to read radiology scans for signs of different medical conditions.
Zebra is best known in the industry for charging $1 for each scan. The founders hoped the simple pricing would make it easier for Zebra’s algorithm to be used more widely. The $1-per-scan has now become focal for all firms. Of course, they have accomplished more than merely set a pricing standard, and that is what will receive attention.
For some years now Zebra has faced a situation common to many startups as they expand. Zebra has access to a cutting-edge technology, but they have had to decide how to apply it to a promising commercial applications. More concretely, Zebra has had to simultaneously find training data, hire in competitive labor markets, and negotiate FDA approvals and other government regulations.
Zebra is located in the entrepreneurial culture of Herzliya, a suburb of Tel Aviv. For its goals that location is both advantage and disadvantage. Here is the upside. The area benefits from nearby universities and a local start-up culture, as well as research offices for large technology companies including Apple and Microsoft.
Israel’s size is also an advantage, believe it or not. Its population is smaller than Sweden and bigger than Denmark. That made it easier to make a deal for data. Zebra has access to anonymized healthcare data through Israel’s taxpayer-funded government healthcare system. It is easier to make that sort of deal in a smaller country, so Zebra got an earlier start than most other firms.
Specifically, there are four Israeli HMOs, and Zebra made an arrangement to use one HMO’s data. Most people in Israel remain in the same HMO throughout their life. Because all health data is digitized and each patient is tracked using an anonymized patient identification number, health information is organized in such a way that it can provide accessible, anonymized medical data to inquiring parties. After coming to terms, Zebra gained access to tens of thousands of x-ray and CT scans.
What is the disadvantage of that location? For one, the tight labor market for talent. In its current position, Zebra has found itself competing with Google, Apple, Microsoft, and startups to hire new machine learning PhDs, who are looking for competitive salaries and interesting, cutting-edge work. To be sure, it has found some. Founded in December 2013, Zebra employed approximately 40 people in 2018, including its three founders, and it has been growing since then.
The second disadvantage is distance from big markets, which are in the U.S., Europe, and India. The founders spend significant time travelling to meet with their partners in hospitals across the globe.
Perhaps most interesting, location does not put Zebra at a technical disadvantage. It uses recent breakthroughs in data storage and computing power by employing Google’s Tensor Flow. That makes data-heavy analysis possible anywhere in the world with access to major internet lines and the technical talent to develop the software. In that sense, Zebra occupies the same technical playing field as everyone else.
What precisely does Zebra’s algorithm do? It relies on images from x-rays and CT scans, which are both radiation-based images that create a view of the internal parts of a body, with shades of gray depicting muscles and fat, and white showing bones and metal. X-rays are the single images from a scan, and CT scans are a series of X-ray scans that together make up a three-dimensional image.
Since the 2000s, these images have been increasingly digitized, which allowed for faster, more efficient, and cheaper methods of analysis. The digital copies allowed researchers to use collections of the scans more easily for research purposes, especially for machine learning purposes. Zebra builds on this digital foundation and the research into how to use it for algorithm development.
To date Zebra’s algorithms use large amounts of data within outlined parameters, i.e., in a framework of supervised learning. Accordingly, each data set addresses one application at a time – i.e., spinal fractures, brain bleeds, lung congestion, and so on. Data is analyzed by a computer program for recognizable patterns, which the program could then recognize in other data sets, or refine when the original set has more data added in.
In Zebra’s case, a series of images have been pre-analyzed, with the sections of the images that showed if there was a potential problem marked and annotated for the computer to learn from. All scanned x-rays for training data must be read by multiple doctors in order to give the software reliable data from which to learn. Scans that are performed with out-of-date equipment, or by technicians who are not trained on the equipment, can lead to difficult-to-analyze results.
That said, X-rays, while a valuable medical tool, are still often unclear or misread. Hence, radiology is not a perfect science. Related, no machine learning algorithm will be perfect.
So what does “great but imperfect” software do? In common parlance, great software reduces “false positives” and “false negatives.” A great test can find nearly all positive diagnoses of a condition. A great test also correctly identify when a person does not have a condition.
Notably, a key aspect of Zebra’s strategy, unlike many other healthcare machine learning developers, is to work closely with healthcare providers to develop its software. Zebra positions its software as a tool that can help radiologists, not replace them. The software for detecting bone fractures, for example, came from conversations with radiologists who explained that a tool could help them screen the elderly with high risk. More broadly, they have learned from that experience that automated diagnosis creates value when the software is a tool for screening, or an instrument to aid decision making triage in urgent care. At some point it may also become a useful tool for “second opinions” or “finding clues humans missed, and developing new concerns to explore.”Zebra still faces the same dilemmas of this business as everyone else: whether Zebra should focus its development on products for the developing world (at first China and India) or focus on the U.S. and Europe. In China and India, Zebra’s diagnostic software can help with medical care in areas that have x-ray and CT scanners, but often not enough trained providers to accurately read the results. These countries also have shorter timelines for government approval, which makes it easier for Zebra to go to market. Conversely, in the U.S. and Europe Zebra’s software will be positioned differently and potentially will be used on every appropriate scan in a busy hospital, allowing Zebra to charge more overall and gather more data to improve its accuracy.
After exploring the potential of all options, to date Zebra has tended towards commercializing in Europe and the US, enduring the regulatory scrutiny of the European Commission and the Federal Drug Administration. Several of its products have already made it through the FDA approval process. By the norms of the field that makes Zebra a leader in commercial applications.
Commercialization is the act of translating technical knowledge into valuable products and services. Zebra has focused most of its energies around commercialization, namely, product development, and the development of efficient processes for developing algorithms for specific diagnoses.
Today they are deploying those processes in specific medical settings. The central issue for management is “which diagnoses should gain their priority and why?
Think of Zebra’s experience as a good barometer of progress in commercialization of machine learning for medical applications. Despite all the excitement, the conclusion is irrefutable: While plenty of value could be created, we are just at the beginning of deployment into mainstream medical practice.
References: Shane Greenstein and Sarah Gulick, “Zebra Medical Vision.” Harvard Business School Case, 619053. November, 2018.
Copyright held by IEEE. To view the original essay click here.
The preceding is republished on TAP with permission by its author, Professor Shane Greenstein, Harvard Business School. “Earning Stripes in Medical Machine Learning” was published on Digitopoly on October 10, 2019. The original article this post is taken from was published with the same title in IEEE Micro September - October 2019 (Vol. 39, Issue 5, Pages 126-128). IEEE holds the copyright.