A tireless diagnostician with great accuracy – machine learning stirs up the healthcare industry

The diagnostic quality in healthcare can be increased with expert systems combining machine vision and machine learning. A well-trained artificial intelligence model can identify patterns in medical images that may be difficult for the human eye to detect. Solutions utilizing machine vision and learning are capable of, for example, disease classification and cancer grading already today.

With an embedded device running an AI model analysing medical images, real-world human expertise can be replicated and distributed globally at an affordable and reliable rate. The traditional way of training more experts and sending them around the world to perform their tasks costs time and money.

Medical institutions and professionals are interested in devices that cost-effectively improve healthcare operations. Therefore, the potential of machine vision and learning for a medtech equipment manufacturer is enormous.

“Any image analysis task is a potential case for machine vision and learning”, says Matthias Zumpe, Etteplan’s Embedded Software Team Manager.
 

An embedded solution already counts the intestinal parasite eggs in a sample. An AI microscope automates imaging and uses artificial intelligence to classify and count different parasites.

Previously human experts with microscopes and counter clickers did the work manually. An embedded device can achieve the same accuracy as a human, only faster and without taking breaks.

“These low-cost devices can be easily delivered to, say, schools in developing countries to help with the diagnostics”, says Zumpe. The devices were developed by Etteplan together with Johnson & Johnson.

Running the machine learning model requires only a modest amount of processing power that can easily be executed locally on the device without a cloud connection.
 

The first step for a healthcare equipment manufacturer that wants to enter the world of AI is to determine what data is available to train the machine learning model on.

“Organisations may sometimes underestimate the amount and quality of data required needed to train a well-performing model”, says Zumpe.

To collect the right sort and amount of data successfully, it is highly recommended to find a partner with experience in machine learning, as it differs significantly from traditional, for example, filter-based, image processing methods.

Developing a product that utilizes AI usually means entering the field of diagnostics and decision making.

“Compared to a more traditional way of simply presenting the images to medical staff to help them make the decisions, interpreting the images and data with a machine learning model comes, of course, with a lot more responsibility and regulations to comply with”, says Zumpe.

In addition, healthcare legislation varies across the globe, which may make developing a compliant solution more complex.
 

As some of the training data for a healthcare application may be sensitive, there may be a need to run the machine learning process on on-site servers.

“Depending on the data, sometimes it is more convenient to use off-the-shelf machine learning services from the public cloud such as Google Cloud, AWS, or Microsoft Azure. We have experience from both environments”, says Zumpe.

A seasoned partner can also help choose the most suitable machine learning model for each case. There certainly is a number of potential pitfalls lurking around that can be hard for a beginner to identify.

“How can you ensure the sensitive patient data does not leak into the model? How do you create non-discriminating, ethical expert systems? This kind of questions are hard to answer without a partner specializing in the machine learning field”, says Zumpe.

Etteplan has experience organizing the whole machine vision and machine learning process from start to end. Of course, the chances are that you may not need machine learning at all.

“We look at every case individually. If the client’s problem can be solved using simpler and more affordable methods than machine learning, we naturally prefer those. Applying machine learning is not an end in itself for us. It is just another tool in our toolbox, though a very powerful one”, says Zumpe.