Machine vision systems involve a combination of software and hardware, including a camera to capture an image and a computer to analyze it with dedicated algorithms. Those algorithms, termed neural networks, interestingly work similarly to the human brain: They find patterns and anomalies in the images and interpret them.
One of the challenges with machine vision is identifying patterns with enough accuracy and confidence to allow the system to make decisions in a real-world situation. We’ve seen this in the first generation of autonomous vehicle systems, where radar was used to provide alerts but was not yet advanced enough to analyze road situations and make driving decisions. It is only the introduction of machine vision technology that enabled self-driving cars.
And this is exactly what is happening now in the field of medical diagnostics, where machine vision-based algorithms enable devices to make deterministic decisions. We should expect to see continued growth, in terms of both adoption and technology, in 2021.
Advancement In AI-Enabled Machine Vision
In pathology, and specifically for complete blood count (CBC), first-generation automation has been applied since the 1960s by running cells through pipes using a technology called flow cytometry. When those devices provide questionable results, the lab technician would manually prepare a blood smear and review thousands of cells under the microscope. This manual review process is expensive and slow.
In the early 2000s, companies started experimenting with machine vision. The first digital microscope systems cost about $300,000 to set up and took over 24 hours to scan a single slide. Naturally, only big central labs with space and dedicated technicians can afford such a system.
Twenty years later, thanks to major technological advancements and hardware improvements, we’re seeing affordable and compact blood analyzers that are making diagnostics more accessible. This automation of the image recognition of tissue or blood samples combines the accuracy of the first-generation technology and the ability of microscopists to analyze abnormalities. This innovation ultimately makes fast and accurate diagnostics accessible in a variety of clinical settings — even those without big and expensive central laboratories.
We see the same breakthrough technologies in other areas such as radiology, where conventionally, physicians visually assessed medical images for the detection, characterization and monitoring of diseases. This manual process often leads to inefficiencies in a time when there are widespread radiologist shortages. With the application of AI methods that excel at recognizing complex patterns in imaging data and providing quantitative assessments of radiographic characteristics, more companies are now working on AI solutions targeted at radiologists to improve the quality and speed of the CT scan process.
Aidoc is one of several healthcare AI companies that provides machine vision image analysis solutions to improve radiologists’ productivity while enhancing patient outcomes. The goal is to optimize radiologists’ workflows, not replace them, by reducing the workload on tedious tasks like segmenting structures to enable more quantitative imaging, thus improving patient care.
Overcoming Challenges
Across all applications of medical AI, an incorrect decision can lead to irreversible consequences since we’re dealing with patients’ lives. Machine learning algorithms are vulnerable to cognitive and technical bias, especially when the image dataset is small or lacking diversity. Effective algorithms must be trained on large datasets and be diverse in race, gender and geography. Without addressing these data needs and how the algorithms are designed, we risk perpetuating inequalities in medicine.
While training an algorithm is limited by the quality of information fed into the system, we should continue to employ a hybrid mix of human review and automated solutions in medical diagnostics. Biology is always complex, and the number of unique conditions is enormous. All healthy cells are alike; each abnormal cell is abnormal in its own way. In pathology, some cells always look different from the “average” cell. In investigating rare phenomena, humans are superior at analyzing things for the first time and are better suited to derive insights through association. The best the computer systems can do in these cases is alerting when they don't recognize a specific image so that the technician can focus only on those cases.
AI software must deal with not only the variability of the external world but also the variability of the hardware of the device. Different devices tend to produce different images — this is because cameras and lenses are all slightly imperfect. Engineers must address this challenge by building high-quality calibration and adjustment processes. Complementary, the AI algorithms should be trained to compensate for this variability. These are the kinds of challenges AI companies are dealing with today, with incredible breakthroughs that will revolutionize the health system.
In the years to come, we’re likely going to see a proliferation of AI-enabled supportive diagnostics systems that offer recommendations to clinicians. However, the real advancement will be evident in the development of new, deterministic systems that will not only support but provide actionable patient results that will improve health outcomes.
About the Author:
Sarah Levy | Forbes Councils Member
Sarah Levy is the CTO of Sight Diagnostics, a company transforming Healthcare through fast, accurate and less painful blood testing
