November 8th. A date not many would remember but 123 years back on this day, Wilhelm Conrad Roentgen made a remarkable discovery. He discovered a process by which one could cast shadows of solid objects on pieces of film by passing rays through it. He called it X-Ray because in mathematics “X” is used to indicate the unknown quantity and he hadn’t quite figured the nature of rays and its value back then.
Fast forward to the 21stCentury. As per P&S Intelligence report“the global x-ray imaging market size is projected to reach $12.4 billion by 2022, growing at a CAGR of 4.9% during the forecast period. The growing geriatric population, increasing prevalence of chronic diseases and increasing need for diagnostic imaging procedures is driving the growth of the market. In addition, technological advancements in X-ray imaging and increasing healthcare expenditure are also bolstering the market growth.”
The scenario seems very promising, especially from a revenue standpoint, but what truly slips through the crack in the rush of optimism is ‘accessibility’.
X-ray machines are proliferating rural areas in several developing countries through a concerted effort of philanthropists, NGOs and the Government, but finding Radiologistsfor analysing all these X-rays is a challenge. Even if a hospital does manage to employ a Radiologist, the workload pressure becomes unmanageable from the huge amount of X-rays generated.
So, quite logically, accessibility is easier said than done.
However, if I were to de-construct the problem and plot it on the dimension of time, I see a huge opportunity to make this work.
Let me illustrate with an example.
The most important component of operational cost is time and to arrive at that cost one needs to map the effort involved in delivering that particular service. The effort involved in the Imaging business can essentially be split in to 2 parts — Imaging and Analysis.
Imaging is fairly simple and effortless. The machine does most of the work and technicians to operate the machines are not too difficult to come by. Analysis, however, requires more effort. There are multiple steps involved in the overall process of analysis- manual screening of the X-Rays, classifying the images across different conditions/disorders, vetting the anomalies, escalation, writing the report, signing off etc.
What’s important to note is that Anomalies can have multiple variations and not every image shows the same pattern. And considering accuracy is essential, Radiologist cannot simply skim through images and use standard markers of anomalies to assign a statement. They have to put a sizeable amount of time to identify issues then painstakingly classify them and then attach a statement to the report.
There’s no short cut to this method. And to be able to deploy a service of this nature at a remote location, more than machines you need trained people, a team of Radiologists to be specific. It’s not cheap to run an Imaging business / service.
However, if we were to create an automated system that could accelerate the process, I reckon there would be a positive impact on efficiency and thereby reduction of cost. Machine Learning (and Deep Learning) have been increasingly considered for this automation.
The biggest obstacle in adoption of Machine Learning is commercial viability. The cost of developing a system that could meet the accuracy levels for a fully automated system (a Radiologist’s need is eliminated completely) is inexplicably prohibitive. Even if the costs were to be managed the challenge that’d be impossible to overcome would be that of required datasets for training. You’d need a critical amount of data both in terms of quality and volume to be able to train the system to reach needed accuracy levels. And chances of procuring such a high volume of X-rays from the same distribution is a difficult task.
We’ve investigated multiple learning algorithms with various pipelines and arrived at this conclusion that it’s next to impossible to build a commercially viable fully automated system with the required accuracy ground-up.
However, not necessarily everything needs to be looked at from the lens of accuracy. We can utilise the power of Machine Learning is various other ways.
Machine Learning can act as the supporting tool and prioritisation as a system can exponentially reduce the cognitive load on Radiologists, enabling them to focus more on clinical decision making.
Prioritisation essentially means an automated system based on Machine Learning that reduces the load of manual screening by offering a prioritised pipeline of pre-classified X-Rays across different related diseases/disorders. A more systematic and structured approach in which Radiologists can work more efficiently as they are presented first with X-rays that are more likely to have a condition requiring attention.
So in a way, if cognitive load can be reduced, the impact on cost will also effectively go down. Which in turn allows services to become relatively viable.
Also, a system like this is not expensive to set up. The process does not require a large set of training data as the need is not for accuracy but efficiency.
We all know that physiological and pathological processes are quite complex and data alone cannot provide a solution. Probabilistic reasoning and clinical inferences combined with the process of elimination are central to decision making.
Radiologists, through years of training and experience can make decisions that the machine after a few days training on a limited data set absolutely cannot. However, a solution that can present the radiologist with a prioritised and classified queue ensuring he/she receives a set with abnormalities first and then a set of normal ones, the overall process becomes more efficient and the load on the manual screening system automatically reduces.
Time equals money.
Over time, the system of prioritisation becomes the foundation for deep learning and can effectively play out the accuracy task as well. Effectively, the prioritisation data can become the base to train the machine for accuracy.
In sum, Machine Learning can democratise opportunities, support viability and promote accessibility, especially in the healthcare industry.
It is to be noted that every Machine Learning project doesn’t necessarily need to start with an objective to achieve output of the highest level. You can look at tasks that can create efficiency to begin with and then use it as a building block to assign complex output tasks such as accuracy or speed.
But the first step is to adopt the idea and then enhance and scale as you go along.