Healthcare is a field that stands to benefit greatly from machine learning on the edge. (Source: ipopba - stock.adobe.com)
Machine learning is one of the most transformative technologies that has emerged in the past decade. As technology continues to grow and evolve, its impact becomes more apparent through its proliferation in and integration into countless fields.
Among these fields, healthcare is particularly important. Today, doctors and medical professionals use machine learning to save and improve lives by detecting and analyzing diseases with accuracies and speeds that would’ve been unfathomable only decades ago. Within this, the field of cancer detection has benefited significantly from the ability to automatically identify and classify tumors.
Despite this success, to fully harness the power of machine learning in healthcare, developers need to bring the technology to the edge—a pursuit that has been hindered by a number of significant challenges. In this blog, we’ll discuss the application of machine learning in healthcare with a focus on cancer detection, why the field benefits from edge computing, and how Edge Impulse has helped make that possible.
At the root of it all, technology’s ultimate goal is to improve people’s lives. Arguably, the healthcare industry is where this altruistic goal is most tangible. Specifically, when you look at the use case of cancer detection, it becomes apparent just how much of a benefit machine learning provides.
Historically, cancer detection has been a meticulous and tedious process that requires the explicit time and attention of a highly trained medical professional. Before anything intrusive is done, the first step in cancer detection is often diagnostic imaging tests, in which a scan is taken of a suspected location. Following the scan, a doctor or radiologist will need to spend their time manually analyzing and interpreting the resulting images to identify any signs of cancer.
All things considered, this process can take a considerable amount of time and may be prone to errors in the human interpretation of images. In a worst-case scenario, this time and lack of accuracy could result in otherwise avoidable deaths.
Here, machine learning has emerged as a revolutionary technology with the potential to save countless lives. Instead of the traditional approach described above, with machine learning, the cancer detection process can remove the medical professional altogether. In this system, images are still taken, but instead of having a medical professional analyze the results, the images can be fed into a machine-learning model trained for tumor and cancer detection.
The benefits of this scheme are significant. First, machine learning-based cancer detection is significantly faster than the alternative. Instead of waiting on the availability of an extremely busy medical professional to analyze the images, a machine-learning model can immediately process the images as soon as they are available. Further, image classification models are extremely accurate, with some studies1 indicating that they are more accurate than human detection. According to Hugo Aerts, Ph.D., of Harvard Medical School, “AI can automate assessments and tasks that humans currently can do but take a lot of time.”
Hence, machine learning can detect cancer faster and more accurately. When people’s lives are at stake, both speed and accuracy are crucial, and the result of machine learning is that more lives can be saved and improved. At the same time, instead of interpreting scans, medical professionals can be freed up to focus on treatment options, further benefiting patients and the healthcare system.
While the potential for machine learning in healthcare is huge, significant shortcomings still exist in the current way the technology is used. Primarily, many of these challenges arise from the fact that, in order to perform machine-learning computation, many healthcare systems rely on cloud computing.
In general, machine learning is an extremely computationally expensive task, requiring dedicated hardware such as GPUs or accelerators to perform with low latency. The challenge is that many healthcare systems don’t have access to such computing resources. Instead, they rely on the cloud. In this scheme, the images generated from the scans are sent wirelessly from the imaging machine to a data center where the computation happens.
One reason that this is a challenge is that it requires a robust and performant network infrastructure to facilitate. Medical scans tend to be extremely detailed and high-resolution, meaning they consist of large amounts of data. Sending this data to and from a data center requires network infrastructures with the bandwidth and performance to handle the generated traffic.
While this may be feasible for well-funded first-world hospitals, the vast majority of medical centers in the world do not have these capabilities. Even further, if we extrapolate this to third-world or remote destinations, wireless connectivity may not be available at all. Hence, the scheme of using cloud computing for healthcare machine learning ends up not being feasible for the majority of hospitals.
Instead, we need to bring healthcare machine learning to the edge to democratize this technology and provide life-saving benefits to people everywhere.
By bringing machine learning to the edge with Edge Impulse, technology experts and companies such as Tiny are making strides toward making healthcare technology more accessible to the masses. Tiny, for example, is engaged in the development of TinyML models to enable healthcare, and specifically cancer detection, on the edge. After considering many different paths toward this goal, Tiny eventually landed on Edge Impulse as the best option.
Traditionally deploying a model to the edge would require engineers to perform every one of the laborious tasks involved in the machine-learning workflow, including collecting and labeling a data set, determining a model architecture, training the model, quantizing it, and generating device-specific binaries for the edge target. From a business perspective, this process is far from ideal as it requires extensive time requirements as well as expertise in frameworks such as PyTorch or TensorFlow.
Instead, the Edge Impulse tool provides a comprehensive and intuitive tool for creating and deploying machine-learning models all the way from data collection to edge deployment. Specifically, Tiny used Edge Impulse for organizing and labeling the collected data as well as training and refining their models until they achieved a desirable accuracy of 90 percent. Within this, Tiny benefited from Edge Impulse’s adaptive processing and fast blocks, making it easy to work with data to get a simple model that can run even on resource-constrained chips. Tiny was then able to automatically generate a device-specific binary to quickly perform machine-learning inference on weak and inexpensive computing devices.
Where the goal of technology is to make the lives of people better, seemingly no industry is more uniquely positioned than healthcare.
With the advancement of technology and the proliferation of machine learning, healthcare has taken great strides, specifically as it pertains to the use case of cancer detection. However, significant technological limitations are still preventing these strides from being enjoyed by all.
With Edge Impulse, Tiny is working on bringing healthcare machine learning to the edge, where the technology can be democratized and shared in all corners of the world. Ultimately, by bringing healthcare machine learning to the edge, technology will be able to improve the healthcare system and even save lives.
Source
Norik Badalyan is an embedded software and embedded machine learning engineer. He currently works at Hi-Tech Gateway as a microelectronics engineer and is the co-founder and engineer for Tiny, an embedded software and embedded machine learning start-up company. Tiny's goal is to create devices for various industries using TinyML and educate students so that TinyML is used more in Armenia and worldwide. In his free time, Norik studies, participates in events and competitions, wins prizes, receives certificates, reads, and writes poetry.