This is part of a series of stories examining how artificial intelligence is disrupting industries.
Can artificial intelligence (AI) make health care smarter? The technology sector is investing heavily in new applications for AI in medicine, in the hope that algorithms can bring new capabilities to medical diagnoses and patient care.
One of the areas where artificial intelligence is emerging as a valuable tool is radiology, which offers an early study in the benefits of AI in medicine, and its potential impact on the healthcare workforce.
Anytime a patient breaks a bone, sprains an ankle or hits their head, the radiology industry goes to work, using x-rays, CT scanners, MRI machines and other tools and techniques to take a closer look inside the human body without the need for surgery. Radiology technicians take the pictures and radiologists examine them to determine the extent of the injury. What if a computer could do the analysis?
The Application of AI in Radiology
AI is being hailed as one of the greatest tools to come to radiology since the x-ray machine, and perhaps also the biggest threat that the industry has faced. AI, or more specifically machine learning and deep learning algorithms, are learning how to analyze the images produced by radiology scans. One case study from 2017 found that with enough data, a deep learning network could determine which patients had breast cancer and which ones were cancer free with 100% accuracy.
In this case study, the AI system was fed 400 high pixel biopsy images to learn to identify different malignancies. These images were 50,000 x 50,000 pixels in scale or more than 520 inches across. Once the system had examined these images and leaned all it could, it was provided with 200 more slides to examine without context. The program was always able to correctly determine which patients had cancer and which did not.
This isn’t the only application for AI in radiology. A team at Colorado State University is using AI and virtual biopsies — without needing to take a tissue sample — to diagnose melanoma. The new technique uses multiphoton microscopy, which provides a dramatic contrast between healthy and malignant tissues.
The NIH recently published a roadmap for artificial intelligence in medical imaging. (Image: National Institutes of Health)
The National Institutes of Health also publicly released a massive dataset of CT images that researchers can use to program and teach AI and machine learning systems. The dataset contains radiology findings to improve lesion identification throughout the body. The 32,000 photos belong to 4,400 anonymized patients. This is the second dataset that NIH has released, the first being a set of chest x-ray images and their data. This move toward collaboration and data sharing could encourage more companies to invest in AI for their radiology departments.
The NIH also has recently published a roadmap for AI in medical imaging examining how standards bodies, professional societies, governmental agencies, and private industry can work together to help patients benefit from the potential of AI to bring about innovative imaging technologies.
The potential use of AI in health care has implications for the data center industry, which could see additional demand for both compute capacity and data storage, given the large size of high-resolution medical images.
Free Resource from Data Center Frontier White Paper Library
Hyperconnected Data Centers: The Key to Digital Transformation
In the era of the hybrid cloud, data and applications are scattered across many locations, including on-premises, colocation facilities, and public cloud services. To enable digital data-driven business strategies, hyperconnected data centers are essential. Download the new webcast from CoreSite, to learn how to implement hyperconnected data centers for digital transformation.Get this PDF emailed to you.
Artificial Intelligence vs. Radiologists
In 2016, Professor Geoffrey Hinton often hailed as the “Godfather of Neural Networks” stated that because of the trend toward artificial intelligence it was “quite obvious that we should stop training radiologists.” In a rather unflattering statement, he described the specialists in this field as “the coyote that’s already over the edge of the cliff who hasn’t yet looked down.”
It’s difficult to predict the future of artificial intelligence in radiology. It will likely become an invaluable tool to assist with diagnosis, but radiologists probably don’t need to worry about Hinton’s cartoon analogy coming to pass. AI medical imagery will likely move out of the labs and into medical facilities where it can change lives, but studies suggest it will never replace radiologists.
A recently published study pitted an AI system against the diagnoses of 101 radiologists. The AI system was programmed to analyze 2,652 exams and rate each patient with a cancer suspicion ranking of 1-10. In spite of a marked performance, the AI system continually performed lower than human radiologists.
Despite this lackluster showing, some believe that AI could potentially replace radiologists, turning the entire industry on its ear – and obviously bad news for those who have made a career out of analyzing medical imagery. Radiology specialists don’t need to start revamping their resumes quite yet, as there are some aspects of the job that no computer will ever replace, from consulting with other physicians to discussing their findings with the patients.
Currently, the existing AI programs for radiology are only used in research settings and are programmed for a very narrow field — looking specifically at breast cancer or melanoma cells. It will be some time before the technology is ready to approach the performance of doctors specializing in radiology, and it will likely never replace radiological technicians, who earn a median income of $56,000 a year.
AI’s most most likely role is as a tool to assist radiologists or specialists. Ultimately, it will always be a human doctor that needs to take responsibility for a diagnosis. An AI system can identify patterns that a human doctor might miss, but the program will never be the one that has to explain a diagnosis to a concerned patient.
Incoming Investments for AI in Radiology
Despite the apparent apprehension surrounding AI in radiology, companies are lining up to invest in this new technology. A medical technology startup in Tel Aviv has raised more than $40 million from investors to develop technology that will use AI to analyze medical images. Instead of requiring a separate computer or server, Aidoc’s software runs on existing hospital hardware, automatically collecting and analyzing any image uploaded to the servers. The company has already received FDA clearance for its AI algorithm that identifies intracranial hemorrhages. The same algorithm is also working in Europe, as well as programs that identify spinal fractures and pulmonary embolism.
Aidoc isn’t the only company trying to get in on the ground floor of AI medical imaging. By 2023, AI in medical imaging will be worth more than $2 billion — nearly four times what it was worth in 2018. ZebraMed is working on creating a massive database of images, slowly releasing automated findings as new discoveries appear. ZebraMed isn’t interested in replacing radiologists — instead, they want to provide them with the tools to augment their practice. This will, ideally, help them identify high-risk patients, optimize workflows, and reduce the cost of care.
VoxelCloud is another company that’s working to make a name for itself in the field of AI medical imaging. The company is using cloud computing to improve medical imaging workflows and is one of the first companies working to provide AI as a service. VoxelCloud is also creating a medical knowledge API where third-party app developers will be able to create their own applications to further augment the field of radiology.
Finally, there is DeepWise. Based in Beijing, this company focuses exclusively on AI and deep learning for medical imagery. There might only be a few companies working toward this end, but the field will fill rapidly as this technology moves out of research labs and into hospitals and medical facilities where they will, hopefully, help save patient lives.
Data Center Frontier is edited by Rich Miller, the data center industry’s most experienced journalist. For more than 15 years, Rich has profiled the key role played by data centers in the Internet revolution.
EXCELLENT ARTICLE ON UP AND COMING COMPANIES !
INFORMATION I CAN USE …….
Hi Kayla,
Your first two examples above are not use cases in Radiology. They both are “Pathology” use cases and not Radiology. The first used pictures taken of biopsy specimens and not x-ray, CT or MRI data. The second also uses photos of the skin and no images from radiology modalities like Xray, CT, Ultrasound or MRI. I think the timing of your article is correct and but the use cases are totally wrong if you are talking about Radiology. I personally think there is huge opportunity for AI in Radiology and we will actually see demand for Radiologists who use AI to make image interpretation and management decisions. Statements like the ones by Geoffrey Hinton are misleading and more of fear tactics and scare medical students from getting into a specialty that already has shortage of physicians. We need to be training our next generation radiologists on how to use AI to improve diagnostic accuracy and better treatment decisions.
Thanks